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) PetscCall(LandauKokkosDestroyMatMaps(maps, maps[0].numgrids)); // implies Kokkos does
34: #endif
35: }
36: // free host data
37: for (PetscInt grid = 0; grid < maps[0].numgrids; grid++) {
38: PetscCall(PetscFree(maps[grid].c_maps));
39: PetscCall(PetscFree(maps[grid].gIdx));
40: }
41: PetscCall(PetscFree(maps));
42: PetscFunctionReturn(PETSC_SUCCESS);
43: }
44: static PetscErrorCode energy_f(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf_dummy, PetscScalar *u, void *actx)
45: {
46: PetscReal v2 = 0;
48: PetscFunctionBegin;
49: /* compute v^2 / 2 */
50: for (PetscInt i = 0; i < dim; ++i) v2 += x[i] * x[i];
51: /* evaluate the Maxwellian */
52: u[0] = v2 / 2;
53: PetscFunctionReturn(PETSC_SUCCESS);
54: }
56: /* needs double */
57: static PetscErrorCode gamma_m1_f(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf_dummy, PetscScalar *u, void *actx)
58: {
59: PetscReal *c2_0_arr = ((PetscReal *)actx);
60: double u2 = 0, c02 = (double)*c2_0_arr, xx;
62: PetscFunctionBegin;
63: /* compute u^2 / 2 */
64: for (PetscInt i = 0; i < dim; ++i) u2 += x[i] * x[i];
65: /* gamma - 1 = g_eps, for conditioning and we only take derivatives */
66: xx = u2 / c02;
67: #if defined(PETSC_USE_DEBUG)
68: u[0] = PetscSqrtReal(1. + xx);
69: #else
70: u[0] = xx / (PetscSqrtReal(1. + xx) + 1.) - 1.; // better conditioned. -1 might help condition and only used for derivative
71: #endif
72: PetscFunctionReturn(PETSC_SUCCESS);
73: }
75: /*
76: LandauFormJacobian_Internal - Evaluates Jacobian matrix.
78: Input Parameters:
79: . globX - input vector
80: . actx - optional user-defined context
81: . dim - dimension
83: Output Parameter:
84: . J0acP - Jacobian matrix filled, not created
85: */
86: static PetscErrorCode LandauFormJacobian_Internal(Vec a_X, Mat JacP, const PetscInt dim, PetscReal shift, void *a_ctx)
87: {
88: LandauCtx *ctx = (LandauCtx *)a_ctx;
89: PetscInt numCells[LANDAU_MAX_GRIDS], Nq, Nb;
90: PetscQuadrature quad;
91: PetscReal Eq_m[LANDAU_MAX_SPECIES]; // could be static data w/o quench (ex2)
92: PetscScalar *cellClosure = NULL;
93: const PetscScalar *xdata = NULL;
94: PetscDS prob;
95: PetscContainer container;
96: P4estVertexMaps *maps;
97: Mat subJ[LANDAU_MAX_GRIDS * LANDAU_MAX_BATCH_SZ];
99: PetscFunctionBegin;
102: PetscAssertPointer(ctx, 5);
103: /* check for matrix container for GPU assembly. Support CPU assembly for debugging */
104: PetscCheck(ctx->plex[0] != NULL, ctx->comm, PETSC_ERR_ARG_WRONG, "Plex not created");
105: PetscCall(PetscLogEventBegin(ctx->events[10], 0, 0, 0, 0));
106: PetscCall(DMGetDS(ctx->plex[0], &prob)); // same DS for all grids
107: PetscCall(PetscObjectQuery((PetscObject)JacP, "assembly_maps", (PetscObject *)&container));
108: if (container) {
109: PetscCheck(ctx->gpu_assembly, ctx->comm, PETSC_ERR_ARG_WRONG, "maps but no GPU assembly");
110: PetscCall(PetscContainerGetPointer(container, (void **)&maps));
111: PetscCheck(maps, ctx->comm, PETSC_ERR_ARG_WRONG, "empty GPU matrix container");
112: for (PetscInt i = 0; i < ctx->num_grids * ctx->batch_sz; i++) subJ[i] = NULL;
113: } else {
114: PetscCheck(!ctx->gpu_assembly, ctx->comm, PETSC_ERR_ARG_WRONG, "No maps but GPU assembly");
115: for (PetscInt tid = 0; tid < ctx->batch_sz; tid++) {
116: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMCreateMatrix(ctx->plex[grid], &subJ[LAND_PACK_IDX(tid, grid)]));
117: }
118: maps = NULL;
119: }
120: // 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)
121: PetscCall(PetscFEGetQuadrature(ctx->fe[0], &quad));
122: PetscCall(PetscQuadratureGetData(quad, NULL, NULL, &Nq, NULL, NULL));
123: PetscCall(PetscFEGetDimension(ctx->fe[0], &Nb));
124: 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);
125: 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);
126: // get metadata for collecting dynamic data
127: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
128: PetscInt cStart, cEnd;
129: PetscCheck(ctx->plex[grid] != NULL, ctx->comm, PETSC_ERR_ARG_WRONG, "Plex not created");
130: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd));
131: numCells[grid] = cEnd - cStart; // grids can have different topology
132: }
133: PetscCall(PetscLogEventEnd(ctx->events[10], 0, 0, 0, 0));
134: 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 */
135: DM pack;
136: PetscCall(VecGetDM(a_X, &pack));
137: PetscCheck(pack, PETSC_COMM_SELF, PETSC_ERR_PLIB, "pack has no DM");
138: PetscCall(PetscLogEventBegin(ctx->events[1], 0, 0, 0, 0));
139: for (PetscInt fieldA = 0; fieldA < ctx->num_species; fieldA++) {
140: Eq_m[fieldA] = ctx->Ez * ctx->t_0 * ctx->charges[fieldA] / (ctx->v_0 * ctx->masses[fieldA]); /* normalize dimensionless */
141: if (dim == 2) Eq_m[fieldA] *= 2 * PETSC_PI; /* add the 2pi term that is not in Landau */
142: }
143: if (!ctx->gpu_assembly) {
144: Vec *locXArray, *globXArray;
145: PetscScalar *cellClosure_it;
146: PetscInt cellClosure_sz = 0, nDMs, Nf[LANDAU_MAX_GRIDS];
147: PetscSection section[LANDAU_MAX_GRIDS], globsection[LANDAU_MAX_GRIDS];
148: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
149: PetscCall(DMGetLocalSection(ctx->plex[grid], §ion[grid]));
150: PetscCall(DMGetGlobalSection(ctx->plex[grid], &globsection[grid]));
151: PetscCall(PetscSectionGetNumFields(section[grid], &Nf[grid]));
152: }
153: /* count cellClosure size */
154: PetscCall(DMCompositeGetNumberDM(pack, &nDMs));
155: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) cellClosure_sz += Nb * Nf[grid] * numCells[grid];
156: PetscCall(PetscMalloc1(cellClosure_sz * ctx->batch_sz, &cellClosure));
157: cellClosure_it = cellClosure;
158: PetscCall(PetscMalloc(sizeof(*locXArray) * nDMs, &locXArray));
159: PetscCall(PetscMalloc(sizeof(*globXArray) * nDMs, &globXArray));
160: PetscCall(DMCompositeGetLocalAccessArray(pack, a_X, nDMs, NULL, locXArray));
161: PetscCall(DMCompositeGetAccessArray(pack, a_X, nDMs, NULL, globXArray));
162: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // OpenMP (once)
163: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
164: Vec locX = locXArray[LAND_PACK_IDX(b_id, grid)], globX = globXArray[LAND_PACK_IDX(b_id, grid)], locX2;
165: PetscInt cStart, cEnd, ei;
166: PetscCall(VecDuplicate(locX, &locX2));
167: PetscCall(DMGlobalToLocalBegin(ctx->plex[grid], globX, INSERT_VALUES, locX2));
168: PetscCall(DMGlobalToLocalEnd(ctx->plex[grid], globX, INSERT_VALUES, locX2));
169: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd));
170: for (ei = cStart; ei < cEnd; ++ei) {
171: PetscScalar *coef = NULL;
172: PetscCall(DMPlexVecGetClosure(ctx->plex[grid], section[grid], locX2, ei, NULL, &coef));
173: PetscCall(PetscMemcpy(cellClosure_it, coef, Nb * Nf[grid] * sizeof(*cellClosure_it))); /* change if LandauIPReal != PetscScalar */
174: PetscCall(DMPlexVecRestoreClosure(ctx->plex[grid], section[grid], locX2, ei, NULL, &coef));
175: cellClosure_it += Nb * Nf[grid];
176: }
177: PetscCall(VecDestroy(&locX2));
178: }
179: }
180: PetscCheck(cellClosure_it - cellClosure == cellClosure_sz * ctx->batch_sz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "iteration wrong %" PetscCount_FMT " != cellClosure_sz = %" PetscInt_FMT, cellClosure_it - cellClosure, cellClosure_sz * ctx->batch_sz);
181: PetscCall(DMCompositeRestoreLocalAccessArray(pack, a_X, nDMs, NULL, locXArray));
182: PetscCall(DMCompositeRestoreAccessArray(pack, a_X, nDMs, NULL, globXArray));
183: PetscCall(PetscFree(locXArray));
184: PetscCall(PetscFree(globXArray));
185: xdata = NULL;
186: } else {
187: PetscMemType mtype;
188: if (ctx->jacobian_field_major_order) { // get data in batch ordering
189: PetscCall(VecScatterBegin(ctx->plex_batch, a_X, ctx->work_vec, INSERT_VALUES, SCATTER_FORWARD));
190: PetscCall(VecScatterEnd(ctx->plex_batch, a_X, ctx->work_vec, INSERT_VALUES, SCATTER_FORWARD));
191: PetscCall(VecGetArrayReadAndMemType(ctx->work_vec, &xdata, &mtype));
192: } else {
193: PetscCall(VecGetArrayReadAndMemType(a_X, &xdata, &mtype));
194: }
195: PetscCheck(mtype == PETSC_MEMTYPE_HOST || ctx->deviceType != LANDAU_CPU, ctx->comm, PETSC_ERR_ARG_WRONG, "CPU run with device data: use -mat_type aij");
196: cellClosure = NULL;
197: }
198: PetscCall(PetscLogEventEnd(ctx->events[1], 0, 0, 0, 0));
199: } else xdata = cellClosure = NULL;
201: /* do it */
202: if (ctx->deviceType == LANDAU_KOKKOS) {
203: #if defined(PETSC_HAVE_KOKKOS)
204: 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));
205: #else
206: SETERRQ(ctx->comm, PETSC_ERR_ARG_WRONG, "-landau_device_type %s not built", "kokkos");
207: #endif
208: } else { /* CPU version */
209: PetscTabulation *Tf; // used for CPU and print info. Same on all grids and all species
210: 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];
211: 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;
212: PetscReal *nu_alpha = (PetscReal *)ctx->SData_d.alpha, *nu_beta = (PetscReal *)ctx->SData_d.beta, *invMass = (PetscReal *)ctx->SData_d.invMass;
213: PetscReal (*lambdas)[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS] = (PetscReal (*)[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS])ctx->SData_d.lambdas;
214: PetscSection section[LANDAU_MAX_GRIDS], globsection[LANDAU_MAX_GRIDS];
215: PetscScalar *coo_vals = NULL;
216: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
217: PetscCall(DMGetLocalSection(ctx->plex[grid], §ion[grid]));
218: PetscCall(DMGetGlobalSection(ctx->plex[grid], &globsection[grid]));
219: PetscCall(PetscSectionGetNumFields(section[grid], &Nf[grid]));
220: }
221: /* count IPf size, etc */
222: PetscCall(PetscDSGetTabulation(prob, &Tf)); // Bf, &Df same for all grids
223: const PetscReal *const BB = Tf[0]->T[0], *const DD = Tf[0]->T[1];
224: ip_offset[0] = ipf_offset[0] = elem_offset[0] = 0;
225: for (PetscInt grid = 0; grid < num_grids; grid++) {
226: PetscInt nfloc = ctx->species_offset[grid + 1] - ctx->species_offset[grid];
227: elem_offset[grid + 1] = elem_offset[grid] + numCells[grid];
228: ip_offset[grid + 1] = ip_offset[grid] + numCells[grid] * Nq;
229: ipf_offset[grid + 1] = ipf_offset[grid] + Nq * nfloc * numCells[grid];
230: }
231: IPf_sz_glb = ipf_offset[num_grids];
232: IPf_sz_tot = IPf_sz_glb * ctx->batch_sz;
233: // prep COO
234: PetscCall(PetscMalloc1(ctx->SData_d.coo_size, &coo_vals)); // allocate every time?
235: if (shift == 0.0) { /* compute dynamic data f and df and init data for Jacobian */
236: #if defined(PETSC_HAVE_THREADSAFETY)
237: double starttime, endtime;
238: starttime = MPI_Wtime();
239: #endif
240: PetscCall(PetscLogEventBegin(ctx->events[8], 0, 0, 0, 0));
241: PetscCall(PetscMalloc4(IPf_sz_tot, &ff, IPf_sz_tot, &dudx, IPf_sz_tot, &dudy, (dim == 3 ? IPf_sz_tot : 0), &dudz));
242: // F df/dx
243: for (PetscInt tid = 0; tid < ctx->batch_sz * elem_offset[num_grids]; tid++) { // for each element
244: 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
245: // find my grid:
246: PetscInt grid = 0;
247: while (b_elem_idx >= elem_offset[grid + 1]) grid++; // yuck search for grid
248: {
249: 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];
250: 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];
251: PetscScalar *coef, coef_buff[LANDAU_MAX_SPECIES * LANDAU_MAX_NQND];
252: PetscReal *invJe = &invJ_a[(ip_offset[grid] + loc_elem * Nq) * dim * dim]; // ingJ is static data on batch 0
253: PetscInt b, f, q;
254: if (cellClosure) {
255: coef = &cellClosure[b_id * IPf_sz_glb + ipf_offset[grid] + loc_elem * Nb * loc_Nf]; // this is const
256: } else {
257: coef = coef_buff;
258: for (f = 0; f < loc_Nf; ++f) {
259: LandauIdx *const Idxs = &maps[grid].gIdx[loc_elem][f][0];
260: for (b = 0; b < Nb; ++b) {
261: PetscInt idx = Idxs[b];
262: if (idx >= 0) {
263: coef[f * Nb + b] = xdata[idx + moffset];
264: } else {
265: idx = -idx - 1;
266: coef[f * Nb + b] = 0;
267: for (q = 0; q < maps[grid].num_face; q++) {
268: PetscInt id = maps[grid].c_maps[idx][q].gid;
269: PetscScalar scale = maps[grid].c_maps[idx][q].scale;
270: coef[f * Nb + b] += scale * xdata[id + moffset];
271: }
272: }
273: }
274: }
275: }
276: /* get f and df */
277: for (PetscInt qi = 0; qi < Nq; qi++) {
278: const PetscReal *invJ = &invJe[qi * dim * dim];
279: const PetscReal *Bq = &BB[qi * Nb];
280: const PetscReal *Dq = &DD[qi * Nb * dim];
281: PetscReal u_x[LANDAU_DIM];
282: /* get f & df */
283: for (f = 0; f < loc_Nf; ++f) {
284: const PetscInt idx = b_id * IPf_sz_glb + ipf_offset[grid] + f * loc_nip + loc_elem * Nq + qi;
285: PetscInt b, e;
286: PetscReal refSpaceDer[LANDAU_DIM];
287: ff[idx] = 0.0;
288: for (PetscInt d = 0; d < LANDAU_DIM; ++d) refSpaceDer[d] = 0.0;
289: for (b = 0; b < Nb; ++b) {
290: const PetscInt cidx = b;
291: ff[idx] += Bq[cidx] * PetscRealPart(coef[f * Nb + cidx]);
292: for (PetscInt d = 0; d < dim; ++d) refSpaceDer[d] += Dq[cidx * dim + d] * PetscRealPart(coef[f * Nb + cidx]);
293: }
294: for (PetscInt d = 0; d < LANDAU_DIM; ++d) {
295: for (e = 0, u_x[d] = 0.0; e < LANDAU_DIM; ++e) u_x[d] += invJ[e * dim + d] * refSpaceDer[e];
296: }
297: dudx[idx] = u_x[0];
298: dudy[idx] = u_x[1];
299: #if LANDAU_DIM == 3
300: dudz[idx] = u_x[2];
301: #endif
302: }
303: } // q
304: } // grid
305: } // grid*batch
306: PetscCall(PetscLogEventEnd(ctx->events[8], 0, 0, 0, 0));
307: #if defined(PETSC_HAVE_THREADSAFETY)
308: endtime = MPI_Wtime();
309: if (ctx->stage) ctx->times[LANDAU_F_DF] += (endtime - starttime);
310: #endif
311: } // Jacobian setup
312: // assemble Jacobian (or mass)
313: for (PetscInt tid = 0; tid < ctx->batch_sz * elem_offset[num_grids]; tid++) { // for each element
314: const PetscInt b_Nelem = elem_offset[num_grids];
315: const PetscInt glb_elem_idx = tid % b_Nelem, b_id = tid / b_Nelem;
316: PetscInt grid = 0;
317: #if defined(PETSC_HAVE_THREADSAFETY)
318: double starttime, endtime;
319: starttime = MPI_Wtime();
320: #endif
321: while (glb_elem_idx >= elem_offset[grid + 1]) grid++;
322: {
323: const PetscInt loc_Nf = ctx->species_offset[grid + 1] - ctx->species_offset[grid], loc_elem = glb_elem_idx - elem_offset[grid];
324: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset), totDim = loc_Nf * Nq, elemMatSize = totDim * totDim;
325: PetscScalar *elemMat;
326: const PetscReal *invJe = &invJ_a[(ip_offset[grid] + loc_elem * Nq) * dim * dim];
327: PetscCall(PetscMalloc1(elemMatSize, &elemMat));
328: PetscCall(PetscMemzero(elemMat, elemMatSize * sizeof(*elemMat)));
329: if (shift == 0.0) { // Jacobian
330: PetscCall(PetscLogEventBegin(ctx->events[4], 0, 0, 0, 0));
331: } else { // mass
332: PetscCall(PetscLogEventBegin(ctx->events[16], 0, 0, 0, 0));
333: }
334: for (PetscInt qj = 0; qj < Nq; ++qj) {
335: const PetscInt jpidx_glb = ip_offset[grid] + qj + loc_elem * Nq;
336: 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
337: PetscInt d, d2, dp, d3, IPf_idx;
338: if (shift == 0.0) { // Jacobian
339: const PetscReal *const invJj = &invJe[qj * dim * dim];
340: 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];
341: const PetscReal vj[3] = {xx[jpidx_glb], yy[jpidx_glb], zz ? zz[jpidx_glb] : 0}, wj = ww[jpidx_glb];
342: // create g2 & g3
343: for (d = 0; d < LANDAU_DIM; d++) { // clear accumulation data D & K
344: gg2_temp[d] = 0;
345: for (d2 = 0; d2 < LANDAU_DIM; d2++) gg3_temp[d][d2] = 0;
346: }
347: /* inner beta reduction */
348: IPf_idx = 0;
349: 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
350: PetscInt nip_loc_r = numCells[grid_r] * Nq, Nfloc_r = Nf[grid_r];
351: for (PetscInt ei_r = 0, loc_fdf_idx = 0; ei_r < numCells[grid_r]; ++ei_r) {
352: for (PetscInt qi = 0; qi < Nq; qi++, ipidx++, loc_fdf_idx++) {
353: const PetscReal wi = ww[ipidx], x = xx[ipidx], y = yy[ipidx];
354: PetscReal temp1[3] = {0, 0, 0}, temp2 = 0;
355: #if LANDAU_DIM == 2
356: 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.;
357: LandauTensor2D(vj, x, y, Ud, Uk, mask);
358: #else
359: 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.;
360: if (ctx->use_relativistic_corrections) {
361: LandauTensor3DRelativistic(vj, x, y, z, U, mask, C_0(ctx->v_0));
362: } else {
363: LandauTensor3D(vj, x, y, z, U, mask);
364: }
365: #endif
366: for (PetscInt f = 0; f < Nfloc_r; ++f) {
367: 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;
368: temp1[0] += dudx[idx] * nu_beta[f + f_off] * invMass[f + f_off] * (*lambdas)[grid][grid_r];
369: temp1[1] += dudy[idx] * nu_beta[f + f_off] * invMass[f + f_off] * (*lambdas)[grid][grid_r];
370: #if LANDAU_DIM == 3
371: temp1[2] += dudz[idx] * nu_beta[f + f_off] * invMass[f + f_off] * (*lambdas)[grid][grid_r];
372: #endif
373: temp2 += ff[idx] * nu_beta[f + f_off] * (*lambdas)[grid][grid_r];
374: }
375: temp1[0] *= wi;
376: temp1[1] *= wi;
377: #if LANDAU_DIM == 3
378: temp1[2] *= wi;
379: #endif
380: temp2 *= wi;
381: #if LANDAU_DIM == 2
382: for (d2 = 0; d2 < 2; d2++) {
383: for (d3 = 0; d3 < 2; ++d3) {
384: /* K = U * grad(f): g2=e: i,A */
385: gg2_temp[d2] += Uk[d2][d3] * temp1[d3];
386: /* D = -U * (I \kron (fx)): g3=f: i,j,A */
387: gg3_temp[d2][d3] += Ud[d2][d3] * temp2;
388: }
389: }
390: #else
391: for (d2 = 0; d2 < 3; ++d2) {
392: for (d3 = 0; d3 < 3; ++d3) {
393: /* K = U * grad(f): g2 = e: i,A */
394: gg2_temp[d2] += U[d2][d3] * temp1[d3];
395: /* D = -U * (I \kron (fx)): g3 = f: i,j,A */
396: gg3_temp[d2][d3] += U[d2][d3] * temp2;
397: }
398: }
399: #endif
400: } // qi
401: } // ei_r
402: IPf_idx += nip_loc_r * Nfloc_r;
403: } /* grid_r - IPs */
404: 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);
405: // add alpha and put in gg2/3
406: for (PetscInt fieldA = 0, f_off = ctx->species_offset[grid]; fieldA < loc_Nf; ++fieldA) {
407: for (d2 = 0; d2 < LANDAU_DIM; d2++) {
408: gg2[fieldA][d2] = gg2_temp[d2] * nu_alpha[fieldA + f_off];
409: for (d3 = 0; d3 < LANDAU_DIM; d3++) gg3[fieldA][d2][d3] = -gg3_temp[d2][d3] * nu_alpha[fieldA + f_off] * invMass[fieldA + f_off];
410: }
411: }
412: /* add electric field term once per IP */
413: for (PetscInt fieldA = 0, f_off = ctx->species_offset[grid]; fieldA < loc_Nf; ++fieldA) gg2[fieldA][LANDAU_DIM - 1] += Eq_m[fieldA + f_off];
414: /* Jacobian transform - g2, g3 */
415: for (PetscInt fieldA = 0; fieldA < loc_Nf; ++fieldA) {
416: for (d = 0; d < dim; ++d) {
417: g2[fieldA][d] = 0.0;
418: for (d2 = 0; d2 < dim; ++d2) {
419: g2[fieldA][d] += invJj[d * dim + d2] * gg2[fieldA][d2];
420: g3[fieldA][d][d2] = 0.0;
421: for (d3 = 0; d3 < dim; ++d3) {
422: for (dp = 0; dp < dim; ++dp) g3[fieldA][d][d2] += invJj[d * dim + d3] * gg3[fieldA][d3][dp] * invJj[d2 * dim + dp];
423: }
424: g3[fieldA][d][d2] *= wj;
425: }
426: g2[fieldA][d] *= wj;
427: }
428: }
429: } else { // mass
430: PetscReal wj = ww[jpidx_glb];
431: /* Jacobian transform - g0 */
432: for (PetscInt fieldA = 0; fieldA < loc_Nf; ++fieldA) {
433: if (dim == 2) {
434: g0[fieldA] = wj * shift * 2. * PETSC_PI; // move this to below and remove g0
435: } else {
436: g0[fieldA] = wj * shift; // move this to below and remove g0
437: }
438: }
439: }
440: /* FE matrix construction */
441: {
442: PetscInt fieldA, d, f, d2, g;
443: const PetscReal *BJq = &BB[qj * Nb], *DIq = &DD[qj * Nb * dim];
444: /* assemble - on the diagonal (I,I) */
445: for (fieldA = 0; fieldA < loc_Nf; fieldA++) {
446: for (f = 0; f < Nb; f++) {
447: const PetscInt i = fieldA * Nb + f; /* Element matrix row */
448: for (g = 0; g < Nb; ++g) {
449: const PetscInt j = fieldA * Nb + g; /* Element matrix column */
450: const PetscInt fOff = i * totDim + j;
451: if (shift == 0.0) {
452: for (d = 0; d < dim; ++d) {
453: elemMat[fOff] += DIq[f * dim + d] * g2[fieldA][d] * BJq[g];
454: for (d2 = 0; d2 < dim; ++d2) elemMat[fOff] += DIq[f * dim + d] * g3[fieldA][d][d2] * DIq[g * dim + d2];
455: }
456: } else { // mass
457: elemMat[fOff] += BJq[f] * g0[fieldA] * BJq[g];
458: }
459: }
460: }
461: }
462: }
463: } /* qj loop */
464: if (shift == 0.0) { // Jacobian
465: PetscCall(PetscLogEventEnd(ctx->events[4], 0, 0, 0, 0));
466: } else {
467: PetscCall(PetscLogEventEnd(ctx->events[16], 0, 0, 0, 0));
468: }
469: #if defined(PETSC_HAVE_THREADSAFETY)
470: endtime = MPI_Wtime();
471: if (ctx->stage) ctx->times[LANDAU_KERNEL] += (endtime - starttime);
472: #endif
473: /* assemble matrix */
474: if (!container) {
475: PetscInt cStart;
476: PetscCall(PetscLogEventBegin(ctx->events[6], 0, 0, 0, 0));
477: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, NULL));
478: PetscCall(DMPlexMatSetClosure(ctx->plex[grid], section[grid], globsection[grid], subJ[LAND_PACK_IDX(b_id, grid)], loc_elem + cStart, elemMat, ADD_VALUES));
479: PetscCall(PetscLogEventEnd(ctx->events[6], 0, 0, 0, 0));
480: } else { // GPU like assembly for debugging
481: 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];
482: 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};
483: 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;
484: /* assemble - from the diagonal (I,I) in this format for DMPlexMatSetClosure */
485: for (fieldA = 0; fieldA < loc_Nf; fieldA++) {
486: LandauIdx *const Idxs = &maps[grid].gIdx[loc_elem][fieldA][0];
487: for (f = 0; f < Nb; f++) {
488: PetscInt idx = Idxs[f];
489: if (idx >= 0) {
490: nr = 1;
491: rows0[0] = idx;
492: row_scale[0] = 1.;
493: } else {
494: idx = -idx - 1;
495: for (q = 0, nr = 0; q < maps[grid].num_face; q++, nr++) {
496: if (maps[grid].c_maps[idx][q].gid < 0) break;
497: rows0[q] = maps[grid].c_maps[idx][q].gid;
498: row_scale[q] = maps[grid].c_maps[idx][q].scale;
499: }
500: }
501: for (g = 0; g < Nb; ++g) {
502: idx = Idxs[g];
503: if (idx >= 0) {
504: nc = 1;
505: cols0[0] = idx;
506: col_scale[0] = 1.;
507: } else {
508: idx = -idx - 1;
509: nc = maps[grid].num_face;
510: for (q = 0, nc = 0; q < maps[grid].num_face; q++, nc++) {
511: if (maps[grid].c_maps[idx][q].gid < 0) break;
512: cols0[q] = maps[grid].c_maps[idx][q].gid;
513: col_scale[q] = maps[grid].c_maps[idx][q].scale;
514: }
515: }
516: const PetscInt i = fieldA * Nb + f; /* Element matrix row */
517: const PetscInt j = fieldA * Nb + g; /* Element matrix column */
518: const PetscScalar Aij = elemMat[i * totDim + j];
519: if (coo_vals) { // mirror (i,j) in CreateStaticGPUData
520: const PetscInt fullNb = coo_elem_fullNb[glb_elem_idx], fullNb2 = fullNb * fullNb;
521: 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];
522: for (PetscInt q = 0, idx2 = idx0; q < nr; q++) {
523: for (PetscInt d = 0; d < nc; d++, idx2++) coo_vals[idx2] = row_scale[q] * col_scale[d] * Aij;
524: }
525: } else {
526: for (q = 0; q < nr; q++) rows[q] = rows0[q] + moffset;
527: for (d = 0; d < nc; d++) cols[d] = cols0[d] + moffset;
528: for (q = 0; q < nr; q++) {
529: for (d = 0; d < nc; d++) vals[q * nc + d] = row_scale[q] * col_scale[d] * Aij;
530: }
531: PetscCall(MatSetValues(JacP, nr, rows, nc, cols, vals, ADD_VALUES));
532: }
533: }
534: }
535: }
536: }
537: if (loc_elem == -1) {
538: PetscCall(PetscPrintf(ctx->comm, "CPU Element matrix\n"));
539: for (PetscInt d = 0; d < totDim; ++d) {
540: for (PetscInt f = 0; f < totDim; ++f) PetscCall(PetscPrintf(ctx->comm, " %12.5e", (double)PetscRealPart(elemMat[d * totDim + f])));
541: PetscCall(PetscPrintf(ctx->comm, "\n"));
542: }
543: exit(12);
544: }
545: PetscCall(PetscFree(elemMat));
546: } /* grid */
547: } /* outer element & batch loop */
548: if (shift == 0.0) { // mass
549: PetscCall(PetscFree4(ff, dudx, dudy, dudz));
550: }
551: if (!container) { // 'CPU' assembly move nest matrix to global JacP
552: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // OpenMP
553: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
554: 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];
555: PetscInt nloc, nzl, colbuf[1024], row;
556: const PetscInt *cols;
557: const PetscScalar *vals;
558: Mat B = subJ[LAND_PACK_IDX(b_id, grid)];
559: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
560: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
561: PetscCall(MatGetSize(B, &nloc, NULL));
562: for (PetscInt i = 0; i < nloc; i++) {
563: PetscCall(MatGetRow(B, i, &nzl, &cols, &vals));
564: PetscCheck(nzl <= 1024, PetscObjectComm((PetscObject)B), PETSC_ERR_PLIB, "Row too big: %" PetscInt_FMT, nzl);
565: for (PetscInt j = 0; j < nzl; j++) colbuf[j] = moffset + cols[j];
566: row = moffset + i;
567: PetscCall(MatSetValues(JacP, 1, &row, nzl, colbuf, vals, ADD_VALUES));
568: PetscCall(MatRestoreRow(B, i, &nzl, &cols, &vals));
569: }
570: PetscCall(MatDestroy(&B));
571: }
572: }
573: }
574: if (coo_vals) {
575: PetscCall(MatSetValuesCOO(JacP, coo_vals, ADD_VALUES));
576: PetscCall(PetscFree(coo_vals));
577: }
578: } /* CPU version */
579: PetscCall(MatAssemblyBegin(JacP, MAT_FINAL_ASSEMBLY));
580: PetscCall(MatAssemblyEnd(JacP, MAT_FINAL_ASSEMBLY));
581: /* clean up */
582: if (cellClosure) PetscCall(PetscFree(cellClosure));
583: if (xdata) PetscCall(VecRestoreArrayReadAndMemType(a_X, &xdata));
584: PetscFunctionReturn(PETSC_SUCCESS);
585: }
587: static PetscErrorCode GeometryDMLandau(DM base, PetscInt point, PetscInt dim, const PetscReal abc[], PetscReal xyz[], void *a_ctx)
588: {
589: PetscReal r = abc[0], z = abc[1];
590: LandauCtx *ctx = (LandauCtx *)a_ctx;
592: PetscFunctionBegin;
593: if (ctx->sphere && dim == 3) { // make sphere: works for one AMR and Q2
594: PetscInt nzero = 0, idx = 0;
595: xyz[0] = r;
596: xyz[1] = z;
597: xyz[2] = abc[2];
598: for (PetscInt i = 0; i < 3; i++) {
599: if (PetscAbs(xyz[i]) < PETSC_SQRT_MACHINE_EPSILON) nzero++;
600: else idx = i;
601: }
602: if (nzero == 2) xyz[idx] *= 1.732050807568877; // sqrt(3)
603: else if (nzero == 1) {
604: for (PetscInt i = 0; i < 3; i++) xyz[i] *= 1.224744871391589; // sqrt(3/2)
605: }
606: } else {
607: xyz[0] = r;
608: xyz[1] = z;
609: if (dim == 3) xyz[2] = abc[2];
610: }
611: PetscFunctionReturn(PETSC_SUCCESS);
612: }
614: /* create DMComposite of meshes for each species group */
615: static PetscErrorCode LandauDMCreateVMeshes(MPI_Comm comm_self, const PetscInt dim, const char prefix[], LandauCtx *ctx, DM pack)
616: {
617: PetscFunctionBegin;
618: { /* p4est, quads */
619: /* Create plex mesh of Landau domain */
620: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
621: PetscReal par_radius = ctx->radius_par[grid], perp_radius = ctx->radius_perp[grid];
622: if (!ctx->sphere && !ctx->simplex) { // 2 or 3D (only 3D option)
623: PetscReal lo[] = {-perp_radius, -par_radius, -par_radius}, hi[] = {perp_radius, par_radius, par_radius};
624: DMBoundaryType periodicity[3] = {DM_BOUNDARY_NONE, dim == 2 ? DM_BOUNDARY_NONE : DM_BOUNDARY_NONE, DM_BOUNDARY_NONE};
625: if (dim == 2) lo[0] = 0;
626: else {
627: lo[1] = -perp_radius;
628: hi[1] = perp_radius; // 3D y is a perp
629: }
630: 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
631: PetscCall(DMLocalizeCoordinates(ctx->plex[grid])); /* needed for periodic */
632: if (dim == 3) PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "cube"));
633: else PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "half-plane"));
634: } else if (dim == 2) {
635: size_t len;
636: PetscCall(PetscStrlen(ctx->filename, &len));
637: if (len) {
638: Vec coords;
639: PetscScalar *x;
640: PetscInt N;
641: char str[] = "-dm_landau_view_file_0";
642: str[21] += grid;
643: PetscCall(DMPlexCreateFromFile(comm_self, ctx->filename, "plexland.c", PETSC_TRUE, &ctx->plex[grid]));
644: PetscCall(DMPlexOrient(ctx->plex[grid]));
645: PetscCall(DMGetCoordinatesLocal(ctx->plex[grid], &coords));
646: PetscCall(VecGetSize(coords, &N));
647: PetscCall(VecGetArray(coords, &x));
648: /* scale by domain size */
649: for (PetscInt i = 0; i < N; i += 2) {
650: x[i + 0] *= ctx->radius_perp[grid];
651: x[i + 1] *= ctx->radius_par[grid];
652: }
653: PetscCall(VecRestoreArray(coords, &x));
654: PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], ctx->filename));
655: PetscCall(PetscInfo(ctx->plex[grid], "%" PetscInt_FMT ") Read %s mesh file (%s)\n", grid, ctx->filename, str));
656: PetscCall(DMViewFromOptions(ctx->plex[grid], NULL, str));
657: } else { // simplex forces a sphere
658: PetscInt numCells = ctx->simplex ? 12 : 6, cell_size = ctx->simplex ? 3 : 4, j;
659: const PetscInt numVerts = 11;
660: PetscInt cellsT[][4] = {
661: {0, 1, 6, 5 },
662: {1, 2, 7, 6 },
663: {2, 3, 8, 7 },
664: {3, 4, 9, 8 },
665: {5, 6, 7, 10},
666: {10, 7, 8, 9 }
667: };
668: PetscInt cellsS[][3] = {
669: {0, 1, 6 },
670: {1, 2, 6 },
671: {6, 2, 7 },
672: {7, 2, 8 },
673: {8, 2, 3 },
674: {8, 3, 4 },
675: {0, 6, 5 },
676: {5, 6, 7 },
677: {5, 7, 10},
678: {10, 7, 9 },
679: {9, 7, 8 },
680: {9, 8, 4 }
681: };
682: const PetscInt *pcell = (const PetscInt *)(ctx->simplex ? &cellsS[0][0] : &cellsT[0][0]);
683: PetscReal coords[11][2], *flatCoords = &coords[0][0];
684: PetscReal rad = ctx->radius[grid];
685: for (j = 0; j < 5; j++) { // outside edge
686: PetscReal z, r, theta = -PETSC_PI / 2 + (j % 5) * PETSC_PI / 4;
687: r = rad * PetscCosReal(theta);
688: coords[j][0] = r;
689: z = rad * PetscSinReal(theta);
690: coords[j][1] = z;
691: }
692: coords[j][0] = 0;
693: coords[j++][1] = -rad * ctx->sphere_inner_radius_90degree[grid];
694: coords[j][0] = rad * ctx->sphere_inner_radius_45degree[grid] * 0.707106781186548;
695: coords[j++][1] = -rad * ctx->sphere_inner_radius_45degree[grid] * 0.707106781186548;
696: coords[j][0] = rad * ctx->sphere_inner_radius_90degree[grid];
697: coords[j++][1] = 0;
698: coords[j][0] = rad * ctx->sphere_inner_radius_45degree[grid] * 0.707106781186548;
699: coords[j++][1] = rad * ctx->sphere_inner_radius_45degree[grid] * 0.707106781186548;
700: coords[j][0] = 0;
701: coords[j++][1] = rad * ctx->sphere_inner_radius_90degree[grid];
702: coords[j][0] = 0;
703: coords[j++][1] = 0;
704: PetscCall(DMPlexCreateFromCellListPetsc(comm_self, 2, numCells, numVerts, cell_size, ctx->interpolate, pcell, 2, flatCoords, &ctx->plex[grid]));
705: PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "semi-circle"));
706: PetscCall(PetscInfo(ctx->plex[grid], "\t%" PetscInt_FMT ") Make circle %s mesh\n", grid, ctx->simplex ? "simplex" : "tensor"));
707: }
708: } else {
709: PetscCheck(dim == 3 && ctx->sphere && !ctx->simplex, ctx->comm, PETSC_ERR_ARG_WRONG, "not: dim == 3 && ctx->sphere && !ctx->simplex");
710: PetscReal rad = ctx->radius[grid] / 1.732050807568877, inner_rad = rad * ctx->sphere_inner_radius_45degree[grid], outer_rad = rad;
711: const PetscInt numCells = 7, cell_size = 8, numVerts = 16;
712: const PetscInt cells[][8] = {
713: {0, 3, 2, 1, 4, 5, 6, 7 },
714: {0, 4, 5, 1, 8, 9, 13, 12},
715: {1, 5, 6, 2, 9, 10, 14, 13},
716: {2, 6, 7, 3, 10, 11, 15, 14},
717: {0, 3, 7, 4, 8, 12, 15, 11},
718: {0, 1, 2, 3, 8, 11, 10, 9 },
719: {4, 7, 6, 5, 12, 13, 14, 15}
720: };
721: PetscReal coords[16 /* numVerts */][3];
722: for (PetscInt j = 0; j < 4; j++) { // inner edge, low
723: coords[j][0] = inner_rad * (j == 0 || j == 3 ? 1 : -1);
724: coords[j][1] = inner_rad * (j / 2 < 1 ? 1 : -1);
725: coords[j][2] = inner_rad * -1;
726: }
727: for (PetscInt j = 0, jj = 4; j < 4; j++, jj++) { // inner edge, hi
728: coords[jj][0] = inner_rad * (j == 0 || j == 3 ? 1 : -1);
729: coords[jj][1] = inner_rad * (j / 2 < 1 ? 1 : -1);
730: coords[jj][2] = inner_rad * 1;
731: }
732: for (PetscInt j = 0, jj = 8; j < 4; j++, jj++) { // outer edge, low
733: coords[jj][0] = outer_rad * (j == 0 || j == 3 ? 1 : -1);
734: coords[jj][1] = outer_rad * (j / 2 < 1 ? 1 : -1);
735: coords[jj][2] = outer_rad * -1;
736: }
737: for (PetscInt j = 0, jj = 12; j < 4; j++, jj++) { // outer edge, hi
738: coords[jj][0] = outer_rad * (j == 0 || j == 3 ? 1 : -1);
739: coords[jj][1] = outer_rad * (j / 2 < 1 ? 1 : -1);
740: coords[jj][2] = outer_rad * 1;
741: }
742: PetscCall(DMPlexCreateFromCellListPetsc(comm_self, 3, numCells, numVerts, cell_size, ctx->interpolate, (const PetscInt *)cells, 3, (const PetscReal *)coords, &ctx->plex[grid]));
743: PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "cubed sphere"));
744: PetscCall(PetscInfo(ctx->plex[grid], "\t%" PetscInt_FMT ") Make cubed sphere %s mesh\n", grid, ctx->simplex ? "simplex" : "tensor"));
745: }
746: PetscCall(DMSetFromOptions(ctx->plex[grid]));
747: } // grid loop
748: PetscCall(PetscObjectSetOptionsPrefix((PetscObject)pack, prefix));
749: { /* convert to p4est (or whatever), wait for discretization to create pack */
750: char convType[256];
751: PetscBool flg;
753: PetscOptionsBegin(ctx->comm, prefix, "Mesh conversion options", "DMPLEX");
754: PetscCall(PetscOptionsFList("-dm_landau_type", "Convert DMPlex to another format (p4est)", "plexland.c", DMList, DMPLEX, convType, 256, &flg));
755: PetscOptionsEnd();
756: if (flg) {
757: ctx->use_p4est = PETSC_TRUE; /* flag for Forest */
758: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
759: DM dmforest;
760: PetscBool isForest;
762: PetscCall(DMConvert(ctx->plex[grid], convType, &dmforest));
763: PetscCheck(dmforest, ctx->comm, PETSC_ERR_PLIB, "Convert failed?");
764: PetscCall(PetscObjectSetOptionsPrefix((PetscObject)dmforest, prefix));
765: PetscCall(DMIsForest(dmforest, &isForest));
766: PetscCheck(isForest, ctx->comm, PETSC_ERR_PLIB, "Converted to non Forest?");
767: if (ctx->sphere) PetscCall(DMForestSetBaseCoordinateMapping(dmforest, GeometryDMLandau, ctx));
768: PetscCall(DMDestroy(&ctx->plex[grid]));
769: ctx->plex[grid] = dmforest; // Forest for adaptivity
770: }
771: } else ctx->use_p4est = PETSC_FALSE; /* flag for Forest */
772: }
773: } /* non-file */
774: PetscCall(DMSetDimension(pack, dim));
775: PetscCall(PetscObjectSetName((PetscObject)pack, "Mesh"));
776: PetscCall(DMSetApplicationContext(pack, ctx));
777: PetscFunctionReturn(PETSC_SUCCESS);
778: }
780: static PetscErrorCode SetupDS(DM pack, PetscInt dim, PetscInt grid, LandauCtx *ctx)
781: {
782: PetscInt ii, i0;
783: char buf[256];
784: PetscSection section;
786: PetscFunctionBegin;
787: for (ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) {
788: if (ii == 0) PetscCall(PetscSNPrintf(buf, sizeof(buf), "e"));
789: else PetscCall(PetscSNPrintf(buf, sizeof(buf), "i%" PetscInt_FMT, ii));
790: /* Setup Discretization - FEM */
791: PetscCall(PetscFECreateDefault(PETSC_COMM_SELF, dim, 1, ctx->simplex, NULL, PETSC_DECIDE, &ctx->fe[ii]));
792: PetscCall(PetscObjectSetName((PetscObject)ctx->fe[ii], buf));
793: PetscCall(DMSetField(ctx->plex[grid], i0, NULL, (PetscObject)ctx->fe[ii]));
794: }
795: PetscCall(DMCreateDS(ctx->plex[grid]));
796: PetscCall(DMGetLocalSection(ctx->plex[grid], §ion));
797: for (PetscInt ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) {
798: if (ii == 0) PetscCall(PetscSNPrintf(buf, sizeof(buf), "se"));
799: else PetscCall(PetscSNPrintf(buf, sizeof(buf), "si%" PetscInt_FMT, ii));
800: PetscCall(PetscSectionSetComponentName(section, i0, 0, buf));
801: }
802: PetscFunctionReturn(PETSC_SUCCESS);
803: }
805: /* Define a Maxwellian function for testing out the operator. */
807: /* Using cartesian velocity space coordinates, the particle */
808: /* density, [1/m^3], is defined according to */
810: /* $$ n=\int_{R^3} dv^3 \left(\frac{m}{2\pi T}\right)^{3/2}\exp [- mv^2/(2T)] $$ */
812: /* Using some constant, c, we normalize the velocity vector into a */
813: /* dimensionless variable according to v=c*x. Thus the density, $n$, becomes */
815: /* $$ n=\int_{R^3} dx^3 \left(\frac{mc^2}{2\pi T}\right)^{3/2}\exp [- mc^2/(2T)*x^2] $$ */
817: /* Defining $\theta=2T/mc^2$, we thus find that the probability density */
818: /* for finding the particle within the interval in a box dx^3 around x is */
820: /* f(x;\theta)=\left(\frac{1}{\pi\theta}\right)^{3/2} \exp [ -x^2/\theta ] */
822: typedef struct {
823: PetscReal v_0;
824: PetscReal kT_m;
825: PetscReal n;
826: PetscReal shift;
827: } MaxwellianCtx;
829: static PetscErrorCode maxwellian(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf_dummy, PetscScalar *u, void *actx)
830: {
831: MaxwellianCtx *mctx = (MaxwellianCtx *)actx;
832: PetscInt i;
833: PetscReal v2 = 0, theta = 2 * mctx->kT_m / (mctx->v_0 * mctx->v_0), shift; /* theta = 2kT/mc^2 */
835: PetscFunctionBegin;
836: /* compute the exponents, v^2 */
837: for (i = 0; i < dim; ++i) v2 += x[i] * x[i];
838: /* evaluate the Maxwellian */
839: if (mctx->shift < 0) shift = -mctx->shift;
840: else {
841: u[0] = mctx->n * PetscPowReal(PETSC_PI * theta, -1.5) * (PetscExpReal(-v2 / theta));
842: shift = mctx->shift;
843: }
844: if (shift != 0.) {
845: v2 = 0;
846: for (i = 0; i < dim - 1; ++i) v2 += x[i] * x[i];
847: v2 += (x[dim - 1] - shift) * (x[dim - 1] - shift);
848: /* evaluate the shifted Maxwellian */
849: u[0] += mctx->n * PetscPowReal(PETSC_PI * theta, -1.5) * (PetscExpReal(-v2 / theta));
850: }
851: PetscFunctionReturn(PETSC_SUCCESS);
852: }
854: /*@
855: DMPlexLandauAddMaxwellians - Add a Maxwellian distribution to a state
857: Collective
859: Input Parameters:
860: + dm - The mesh (local)
861: . time - Current time
862: . temps - Temperatures of each species (global)
863: . ns - Number density of each species (global)
864: . grid - index into current grid - just used for offset into `temp` and `ns`
865: . b_id - batch index
866: . n_batch - number of batches
867: - actx - Landau context
869: Output Parameter:
870: . X - The state (local to this grid)
872: Level: beginner
874: .seealso: `DMPlexLandauCreateVelocitySpace()`
875: @*/
876: PetscErrorCode DMPlexLandauAddMaxwellians(DM dm, Vec X, PetscReal time, PetscReal temps[], PetscReal ns[], PetscInt grid, PetscInt b_id, PetscInt n_batch, void *actx)
877: {
878: LandauCtx *ctx = (LandauCtx *)actx;
879: PetscErrorCode (*initu[LANDAU_MAX_SPECIES])(PetscInt, PetscReal, const PetscReal[], PetscInt, PetscScalar[], void *);
880: PetscInt dim;
881: MaxwellianCtx *mctxs[LANDAU_MAX_SPECIES], data[LANDAU_MAX_SPECIES];
883: PetscFunctionBegin;
884: PetscCall(DMGetDimension(dm, &dim));
885: if (!ctx) PetscCall(DMGetApplicationContext(dm, &ctx));
886: for (PetscInt ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) {
887: mctxs[i0] = &data[i0];
888: data[i0].v_0 = ctx->v_0; // v_0 same for all grids
889: data[i0].kT_m = ctx->k * temps[ii] / ctx->masses[ii]; /* kT/m */
890: data[i0].n = ns[ii];
891: initu[i0] = maxwellian;
892: data[i0].shift = 0;
893: }
894: data[0].shift = ctx->electronShift;
895: /* need to make ADD_ALL_VALUES work - TODO */
896: PetscCall(DMProjectFunction(dm, time, initu, (void **)mctxs, INSERT_ALL_VALUES, X));
897: PetscFunctionReturn(PETSC_SUCCESS);
898: }
900: /*
901: LandauSetInitialCondition - Adds Maxwellians with context
903: Collective
905: Input Parameters:
906: . dm - The mesh
907: - grid - index into current grid - just used for offset into temp and ns
908: . b_id - batch index
909: - n_batch - number of batches
910: + actx - Landau context with T and n
912: Output Parameter:
913: . X - The state
915: Level: beginner
917: .seealso: `DMPlexLandauCreateVelocitySpace()`, `DMPlexLandauAddMaxwellians()`
918: */
919: static PetscErrorCode LandauSetInitialCondition(DM dm, Vec X, PetscInt grid, PetscInt b_id, PetscInt n_batch, void *actx)
920: {
921: LandauCtx *ctx = (LandauCtx *)actx;
923: PetscFunctionBegin;
924: if (!ctx) PetscCall(DMGetApplicationContext(dm, &ctx));
925: PetscCall(VecZeroEntries(X));
926: PetscCall(DMPlexLandauAddMaxwellians(dm, X, 0.0, ctx->thermal_temps, ctx->n, grid, b_id, n_batch, ctx));
927: PetscFunctionReturn(PETSC_SUCCESS);
928: }
930: // adapt a level once. Forest in/out
931: #if defined(PETSC_USE_INFO)
932: static const char *s_refine_names[] = {"RE", "Z1", "Origin", "Z2", "Uniform"};
933: #endif
934: static PetscErrorCode adaptToleranceFEM(PetscFE fem, Vec sol, PetscInt type, PetscInt grid, LandauCtx *ctx, DM *newForest)
935: {
936: DM forest, plex, adaptedDM = NULL;
937: PetscDS prob;
938: PetscBool isForest;
939: PetscQuadrature quad;
940: PetscInt Nq, Nb, *Nb2, cStart, cEnd, c, dim, qj, k;
941: DMLabel adaptLabel = NULL;
943: PetscFunctionBegin;
944: forest = ctx->plex[grid];
945: PetscCall(DMCreateDS(forest));
946: PetscCall(DMGetDS(forest, &prob));
947: PetscCall(DMGetDimension(forest, &dim));
948: PetscCall(DMIsForest(forest, &isForest));
949: PetscCheck(isForest, ctx->comm, PETSC_ERR_ARG_WRONG, "! Forest");
950: PetscCall(DMConvert(forest, DMPLEX, &plex));
951: PetscCall(DMPlexGetHeightStratum(plex, 0, &cStart, &cEnd));
952: PetscCall(DMLabelCreate(PETSC_COMM_SELF, "adapt", &adaptLabel));
953: PetscCall(PetscFEGetQuadrature(fem, &quad));
954: PetscCall(PetscQuadratureGetData(quad, NULL, NULL, &Nq, NULL, NULL));
955: 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);
956: PetscCall(PetscFEGetDimension(ctx->fe[0], &Nb));
957: PetscCall(PetscDSGetDimensions(prob, &Nb2));
958: PetscCheck(Nb2[0] == Nb, ctx->comm, PETSC_ERR_ARG_WRONG, " Nb = %" PetscInt_FMT " != Nb (%" PetscInt_FMT ")", Nb, Nb2[0]);
959: 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);
960: PetscCall(PetscInfo(sol, "%" PetscInt_FMT ") Refine phase: %s\n", grid, s_refine_names[type]));
961: if (type == 4) {
962: for (c = cStart; c < cEnd; c++) PetscCall(DMLabelSetValue(adaptLabel, c, DM_ADAPT_REFINE));
963: } else if (type == 2) {
964: PetscInt rCellIdx[8], nr = 0, nrmax = (dim == 3) ? 8 : 2;
965: PetscReal minRad = PETSC_INFINITY, r;
966: for (c = cStart; c < cEnd; c++) {
967: PetscReal tt, v0[LANDAU_MAX_NQND * 3], J[LANDAU_MAX_NQND * 9], invJ[LANDAU_MAX_NQND * 9], detJ[LANDAU_MAX_NQND];
968: PetscCall(DMPlexComputeCellGeometryFEM(plex, c, quad, v0, J, invJ, detJ));
969: (void)J;
970: (void)invJ;
971: for (qj = 0; qj < Nq; ++qj) {
972: tt = PetscSqr(v0[dim * qj + 0]) + PetscSqr(v0[dim * qj + 1]) + PetscSqr((dim == 3) ? v0[dim * qj + 2] : 0);
973: r = PetscSqrtReal(tt);
974: if (r < minRad - PETSC_SQRT_MACHINE_EPSILON * 10.) {
975: minRad = r;
976: nr = 0;
977: rCellIdx[nr++] = c;
978: 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));
979: } else if ((r - minRad) < PETSC_SQRT_MACHINE_EPSILON * 100. && nr < nrmax) {
980: for (k = 0; k < nr; k++)
981: if (c == rCellIdx[k]) break;
982: if (k == nr) {
983: rCellIdx[nr++] = c;
984: 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)));
985: }
986: }
987: }
988: }
989: for (k = 0; k < nr; k++) PetscCall(DMLabelSetValue(adaptLabel, rCellIdx[k], DM_ADAPT_REFINE));
990: 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));
991: } else if (type == 0 || type == 1 || type == 3) { /* refine along r=0 axis */
992: PetscScalar *coef = NULL;
993: Vec coords;
994: PetscInt csize, Nv, d, nz, nrefined = 0;
995: DM cdm;
996: PetscSection cs;
997: PetscCall(DMGetCoordinatesLocal(forest, &coords));
998: PetscCall(DMGetCoordinateDM(forest, &cdm));
999: PetscCall(DMGetLocalSection(cdm, &cs));
1000: for (c = cStart; c < cEnd; c++) {
1001: PetscInt doit = 0, outside = 0;
1002: PetscCall(DMPlexVecGetClosure(cdm, cs, coords, c, &csize, &coef));
1003: Nv = csize / dim;
1004: for (nz = d = 0; d < Nv; d++) {
1005: PetscReal z = PetscRealPart(coef[d * dim + (dim - 1)]), x = PetscSqr(PetscRealPart(coef[d * dim + 0])) + ((dim == 3) ? PetscSqr(PetscRealPart(coef[d * dim + 1])) : 0);
1006: x = PetscSqrtReal(x);
1007: if (type == 0) {
1008: 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 */
1009: } else if (type == 1 && (z > ctx->vperp0_radius1 || z < -ctx->vperp0_radius1)) {
1010: outside++; /* don't refine outside electron refine radius */
1011: PetscCall(PetscInfo(sol, "\t%" PetscInt_FMT ") (debug) found %s cells\n", grid, s_refine_names[type]));
1012: } else if (type == 3 && (z > ctx->vperp0_radius2 || z < -ctx->vperp0_radius2)) {
1013: outside++; /* refine r=0 cells on refinement front */
1014: PetscCall(PetscInfo(sol, "\t%" PetscInt_FMT ") (debug) found %s cells\n", grid, s_refine_names[type]));
1015: }
1016: if (x < PETSC_MACHINE_EPSILON * 10. && (type != 0 || ctx->re_radius > PETSC_SQRT_MACHINE_EPSILON)) nz++;
1017: }
1018: PetscCall(DMPlexVecRestoreClosure(cdm, cs, coords, c, &csize, &coef));
1019: if (doit || (outside < Nv && nz)) {
1020: PetscCall(DMLabelSetValue(adaptLabel, c, DM_ADAPT_REFINE));
1021: nrefined++;
1022: }
1023: }
1024: PetscCall(PetscInfo(sol, "\t%" PetscInt_FMT ") Refined %" PetscInt_FMT " cells\n", grid, nrefined));
1025: }
1026: PetscCall(DMDestroy(&plex));
1027: PetscCall(DMAdaptLabel(forest, adaptLabel, &adaptedDM));
1028: PetscCall(DMLabelDestroy(&adaptLabel));
1029: *newForest = adaptedDM;
1030: if (adaptedDM) {
1031: if (isForest) PetscCall(DMForestSetAdaptivityForest(adaptedDM, NULL)); // ????
1032: PetscCall(DMConvert(adaptedDM, DMPLEX, &plex));
1033: PetscCall(DMPlexGetHeightStratum(plex, 0, &cStart, &cEnd));
1034: PetscCall(PetscInfo(sol, "\t\t\t\t%" PetscInt_FMT ") %" PetscInt_FMT " cells, %" PetscInt_FMT " total quadrature points\n", grid, cEnd - cStart, Nq * (cEnd - cStart)));
1035: PetscCall(DMDestroy(&plex));
1036: } else *newForest = NULL;
1037: PetscFunctionReturn(PETSC_SUCCESS);
1038: }
1040: // forest goes in (ctx->plex[grid]), plex comes out
1041: static PetscErrorCode adapt(PetscInt grid, LandauCtx *ctx, Vec *uu)
1042: {
1043: PetscInt adaptIter;
1045: PetscFunctionBegin;
1046: 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]};
1047: for (type = 0; type < 5; type++) {
1048: for (adaptIter = 0; adaptIter < limits[type]; adaptIter++) {
1049: DM newForest = NULL;
1050: PetscCall(adaptToleranceFEM(ctx->fe[0], *uu, type, grid, ctx, &newForest));
1051: if (newForest) {
1052: PetscCall(DMDestroy(&ctx->plex[grid]));
1053: PetscCall(VecDestroy(uu));
1054: PetscCall(DMCreateGlobalVector(newForest, uu));
1055: PetscCall(LandauSetInitialCondition(newForest, *uu, grid, 0, 1, ctx));
1056: ctx->plex[grid] = newForest;
1057: } else {
1058: PetscCall(PetscInfo(*uu, "No refinement\n"));
1059: }
1060: }
1061: }
1062: PetscCall(PetscObjectSetName((PetscObject)*uu, "uAMR"));
1063: PetscFunctionReturn(PETSC_SUCCESS);
1064: }
1066: // make log(Lambdas) from NRL Plasma formulary
1067: static PetscErrorCode makeLambdas(LandauCtx *ctx)
1068: {
1069: PetscFunctionBegin;
1070: for (PetscInt gridi = 0; gridi < ctx->num_grids; gridi++) {
1071: PetscInt iii = ctx->species_offset[gridi];
1072: PetscReal Ti_ev = (ctx->thermal_temps[iii] / 1.1604525e7) * 1000; // convert (back) to eV
1073: PetscReal ni = ctx->n[iii] * ctx->n_0;
1074: for (PetscInt gridj = gridi; gridj < ctx->num_grids; gridj++) {
1075: PetscInt jjj = ctx->species_offset[gridj];
1076: PetscReal Zj = ctx->charges[jjj] / 1.6022e-19;
1077: if (gridi == 0) {
1078: if (gridj == 0) { // lam_ee
1079: ctx->lambdas[gridi][gridj] = 23.5 - PetscLogReal(PetscSqrtReal(ni) * PetscPowReal(Ti_ev, -1.25)) - PetscSqrtReal(1e-5 + PetscSqr(PetscLogReal(Ti_ev) - 2) / 16);
1080: } else { // lam_ei == lam_ie
1081: if (10 * Zj * Zj > Ti_ev) {
1082: ctx->lambdas[gridi][gridj] = ctx->lambdas[gridj][gridi] = 23 - PetscLogReal(PetscSqrtReal(ni) * Zj * PetscPowReal(Ti_ev, -1.5));
1083: } else {
1084: ctx->lambdas[gridi][gridj] = ctx->lambdas[gridj][gridi] = 24 - PetscLogReal(PetscSqrtReal(ni) / Ti_ev);
1085: }
1086: }
1087: } else { // lam_ii'
1088: PetscReal mui = ctx->masses[iii] / 1.6720e-27, Zi = ctx->charges[iii] / 1.6022e-19;
1089: PetscReal Tj_ev = (ctx->thermal_temps[jjj] / 1.1604525e7) * 1000; // convert (back) to eV
1090: PetscReal muj = ctx->masses[jjj] / 1.6720e-27;
1091: PetscReal nj = ctx->n[jjj] * ctx->n_0;
1092: 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));
1093: }
1094: }
1095: }
1096: //PetscReal v0 = PetscSqrtReal(ctx->k * ctx->thermal_temps[iii] / ctx->masses[iii]); /* arbitrary units for non-dimensionalization: plasma formulary def */
1097: PetscFunctionReturn(PETSC_SUCCESS);
1098: }
1100: static PetscErrorCode ProcessOptions(LandauCtx *ctx, const char prefix[])
1101: {
1102: PetscBool flg, fileflg;
1103: PetscInt ii, nt, nm, nc, num_species_grid[LANDAU_MAX_GRIDS], non_dim_grid;
1104: PetscReal lnLam = 10;
1105: DM dummy;
1107: PetscFunctionBegin;
1108: PetscCall(DMCreate(ctx->comm, &dummy));
1109: /* get options - initialize context */
1110: ctx->verbose = 1; // should be 0 for silent compliance
1111: ctx->batch_sz = 1;
1112: ctx->batch_view_idx = 0;
1113: ctx->interpolate = PETSC_TRUE;
1114: ctx->gpu_assembly = PETSC_TRUE;
1115: ctx->norm_state = 0;
1116: ctx->electronShift = 0;
1117: ctx->M = NULL;
1118: ctx->J = NULL;
1119: /* geometry and grids */
1120: ctx->sphere = PETSC_FALSE;
1121: ctx->use_p4est = PETSC_FALSE;
1122: ctx->simplex = PETSC_FALSE;
1123: for (PetscInt grid = 0; grid < LANDAU_MAX_GRIDS; grid++) {
1124: ctx->radius[grid] = 5.; /* thermal radius (velocity) */
1125: ctx->radius_perp[grid] = 5.; /* thermal radius (velocity) */
1126: ctx->radius_par[grid] = 5.; /* thermal radius (velocity) */
1127: ctx->numAMRRefine[grid] = 0;
1128: ctx->postAMRRefine[grid] = 0;
1129: ctx->species_offset[grid + 1] = 1; // one species default
1130: num_species_grid[grid] = 0;
1131: ctx->plex[grid] = NULL; /* cache as expensive to Convert */
1132: }
1133: ctx->species_offset[0] = 0;
1134: ctx->re_radius = 0.;
1135: ctx->vperp0_radius1 = 0;
1136: ctx->vperp0_radius2 = 0;
1137: ctx->nZRefine1 = 0;
1138: ctx->nZRefine2 = 0;
1139: ctx->numRERefine = 0;
1140: num_species_grid[0] = 1; // one species default
1141: /* species - [0] electrons, [1] one ion species eg, duetarium, [2] heavy impurity ion, ... */
1142: ctx->charges[0] = -1; /* electron charge (MKS) */
1143: ctx->masses[0] = 1 / 1835.469965278441013; /* temporary value in proton mass */
1144: ctx->n[0] = 1;
1145: ctx->v_0 = 1; /* thermal velocity, we could start with a scale != 1 */
1146: ctx->thermal_temps[0] = 1;
1147: /* constants, etc. */
1148: ctx->epsilon0 = 8.8542e-12; /* permittivity of free space (MKS) F/m */
1149: ctx->k = 1.38064852e-23; /* Boltzmann constant (MKS) J/K */
1150: ctx->n_0 = 1.e20; /* typical plasma n, but could set it to 1 */
1151: ctx->Ez = 0;
1152: for (PetscInt grid = 0; grid < LANDAU_NUM_TIMERS; grid++) ctx->times[grid] = 0;
1153: for (PetscInt ii = 0; ii < LANDAU_DIM; ii++) ctx->cells0[ii] = 2;
1154: if (LANDAU_DIM == 2) ctx->cells0[0] = 1;
1155: ctx->use_matrix_mass = PETSC_FALSE;
1156: ctx->use_relativistic_corrections = PETSC_FALSE;
1157: ctx->use_energy_tensor_trick = PETSC_FALSE; /* Use Eero's trick for energy conservation v --> grad(v^2/2) */
1158: ctx->SData_d.w = NULL;
1159: ctx->SData_d.x = NULL;
1160: ctx->SData_d.y = NULL;
1161: ctx->SData_d.z = NULL;
1162: ctx->SData_d.invJ = NULL;
1163: ctx->jacobian_field_major_order = PETSC_FALSE;
1164: ctx->SData_d.coo_elem_offsets = NULL;
1165: ctx->SData_d.coo_elem_point_offsets = NULL;
1166: ctx->SData_d.coo_elem_fullNb = NULL;
1167: ctx->SData_d.coo_size = 0;
1168: PetscOptionsBegin(ctx->comm, prefix, "Options for Fokker-Plank-Landau collision operator", "none");
1169: #if defined(PETSC_HAVE_KOKKOS)
1170: ctx->deviceType = LANDAU_KOKKOS;
1171: PetscCall(PetscStrncpy(ctx->filename, "kokkos", sizeof(ctx->filename)));
1172: #else
1173: ctx->deviceType = LANDAU_CPU;
1174: PetscCall(PetscStrncpy(ctx->filename, "cpu", sizeof(ctx->filename)));
1175: #endif
1176: PetscCall(PetscOptionsString("-dm_landau_device_type", "Use kernels on 'cpu' 'kokkos'", "plexland.c", ctx->filename, ctx->filename, sizeof(ctx->filename), NULL));
1177: PetscCall(PetscStrcmp("cpu", ctx->filename, &flg));
1178: if (flg) {
1179: ctx->deviceType = LANDAU_CPU;
1180: } else {
1181: PetscCall(PetscStrcmp("kokkos", ctx->filename, &flg));
1182: PetscCheck(flg, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_device_type %s", ctx->filename);
1183: ctx->deviceType = LANDAU_KOKKOS;
1184: }
1185: ctx->filename[0] = '\0';
1186: PetscCall(PetscOptionsString("-dm_landau_filename", "file to read mesh from", "plexland.c", ctx->filename, ctx->filename, sizeof(ctx->filename), &fileflg));
1187: PetscCall(PetscOptionsReal("-dm_landau_electron_shift", "Shift in thermal velocity of electrons", "none", ctx->electronShift, &ctx->electronShift, NULL));
1188: PetscCall(PetscOptionsInt("-dm_landau_verbose", "Level of verbosity output", "plexland.c", ctx->verbose, &ctx->verbose, NULL));
1189: PetscCall(PetscOptionsInt("-dm_landau_batch_size", "Number of 'vertices' to batch", "ex2.c", ctx->batch_sz, &ctx->batch_sz, NULL));
1190: PetscCheck(LANDAU_MAX_BATCH_SZ >= ctx->batch_sz, ctx->comm, PETSC_ERR_ARG_WRONG, "LANDAU_MAX_BATCH_SZ %d < ctx->batch_sz %" PetscInt_FMT, LANDAU_MAX_BATCH_SZ, ctx->batch_sz);
1191: 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));
1192: 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);
1193: PetscCall(PetscOptionsReal("-dm_landau_Ez", "Initial parallel electric field in unites of Conner-Hastie critical field", "plexland.c", ctx->Ez, &ctx->Ez, NULL));
1194: PetscCall(PetscOptionsReal("-dm_landau_n_0", "Normalization constant for number density", "plexland.c", ctx->n_0, &ctx->n_0, NULL));
1195: 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));
1196: PetscCall(PetscOptionsBool("-dm_landau_use_relativistic_corrections", "Use relativistic corrections", "plexland.c", ctx->use_relativistic_corrections, &ctx->use_relativistic_corrections, NULL));
1197: PetscCall(PetscOptionsBool("-dm_landau_simplex", "Use simplex elements", "plexland.c", ctx->simplex, &ctx->simplex, NULL));
1198: PetscCall(PetscOptionsBool("-dm_landau_sphere", "use sphere/semi-circle domain instead of rectangle", "plexland.c", ctx->sphere, &ctx->sphere, NULL));
1199: if (LANDAU_DIM == 2 && ctx->use_relativistic_corrections) ctx->use_relativistic_corrections = PETSC_FALSE; // should warn
1200: 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,
1201: &ctx->use_energy_tensor_trick, NULL));
1203: /* get num species with temperature, set defaults */
1204: for (ii = 1; ii < LANDAU_MAX_SPECIES; ii++) {
1205: ctx->thermal_temps[ii] = 1;
1206: ctx->charges[ii] = 1;
1207: ctx->masses[ii] = 1;
1208: ctx->n[ii] = 1;
1209: }
1210: nt = LANDAU_MAX_SPECIES;
1211: 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));
1212: PetscCheck(flg, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_thermal_temps ,t1,t2,.. must be provided to set the number of species");
1213: PetscCall(PetscInfo(dummy, "num_species set to number of thermal temps provided (%" PetscInt_FMT ")\n", nt));
1214: ctx->num_species = nt;
1215: for (ii = 0; ii < ctx->num_species; ii++) ctx->thermal_temps[ii] *= 1.1604525e7; /* convert to Kelvin */
1216: nm = LANDAU_MAX_SPECIES - 1;
1217: 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));
1218: 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);
1219: nm = LANDAU_MAX_SPECIES;
1220: PetscCall(PetscOptionsRealArray("-dm_landau_n", "Number density of each species = n_s * n_0", "plexland.c", ctx->n, &nm, &flg));
1221: 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);
1222: for (ii = 0; ii < LANDAU_MAX_SPECIES; ii++) ctx->masses[ii] *= 1.6720e-27; /* scale by proton mass kg */
1223: ctx->masses[0] = 9.10938356e-31; /* electron mass kg (should be about right already) */
1224: nc = LANDAU_MAX_SPECIES - 1;
1225: 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));
1226: 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);
1227: for (ii = 0; ii < LANDAU_MAX_SPECIES; ii++) ctx->charges[ii] *= 1.6022e-19; /* electron/proton charge (MKS) */
1228: /* geometry and grids */
1229: nt = LANDAU_MAX_GRIDS;
1230: 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));
1231: if (flg) {
1232: ctx->num_grids = nt;
1233: for (ii = nt = 0; ii < ctx->num_grids; ii++) nt += num_species_grid[ii];
1234: 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,
1235: ctx->num_grids, LANDAU_MAX_GRIDS);
1236: } else {
1237: if (ctx->num_species > LANDAU_MAX_GRIDS) {
1238: num_species_grid[0] = 1;
1239: num_species_grid[1] = ctx->num_species - 1;
1240: ctx->num_grids = 2;
1241: } else {
1242: ctx->num_grids = ctx->num_species;
1243: for (ii = 0; ii < ctx->num_grids; ii++) num_species_grid[ii] = 1;
1244: }
1245: }
1246: 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];
1247: 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],
1248: ctx->num_species);
1249: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1250: PetscInt iii = ctx->species_offset[grid]; // normalize with first (arbitrary) species on grid
1251: ctx->thermal_speed[grid] = PetscSqrtReal(ctx->k * ctx->thermal_temps[iii] / ctx->masses[iii]); /* arbitrary units for non-dimensionalization: plasma formulary def */
1252: }
1253: // get lambdas here because we need them for t_0 etc
1254: PetscCall(PetscOptionsReal("-dm_landau_ln_lambda", "Universal cross section parameter. Default uses NRL formulas", "plexland.c", lnLam, &lnLam, &flg));
1255: if (flg) {
1256: for (PetscInt grid = 0; grid < LANDAU_MAX_GRIDS; grid++) {
1257: for (PetscInt gridj = 0; gridj < LANDAU_MAX_GRIDS; gridj++) ctx->lambdas[gridj][grid] = lnLam; /* cross section ratio large - small angle collisions */
1258: }
1259: } else {
1260: PetscCall(makeLambdas(ctx));
1261: }
1262: non_dim_grid = 0;
1263: 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));
1264: if (non_dim_grid != 0) PetscCall(PetscInfo(dummy, "Normalization grid set to %" PetscInt_FMT ", but non-default not well verified\n", non_dim_grid));
1265: 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);
1266: ctx->v_0 = ctx->thermal_speed[non_dim_grid]; /* arbitrary units for non dimensionalization: global mean velocity in 1D of electrons */
1267: ctx->m_0 = ctx->masses[non_dim_grid]; /* arbitrary reference mass, electrons */
1268: 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 */
1269: /* domain */
1270: nt = LANDAU_MAX_GRIDS;
1271: PetscCall(PetscOptionsRealArray("-dm_landau_domain_radius", "Phase space size in units of thermal velocity of grid", "plexland.c", ctx->radius, &nt, &flg));
1272: if (flg) {
1273: 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);
1274: while (nt--) ctx->radius_par[nt] = ctx->radius_perp[nt] = ctx->radius[nt];
1275: } else {
1276: nt = LANDAU_MAX_GRIDS;
1277: 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));
1278: 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);
1279: 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));
1280: 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);
1281: }
1282: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1283: if (flg && ctx->radius[grid] <= 0) { /* negative is ratio of c - need to set par and perp with this -- todo */
1284: if (ctx->radius[grid] == 0) ctx->radius[grid] = 0.75;
1285: else ctx->radius[grid] = -ctx->radius[grid];
1286: 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)
1287: PetscCall(PetscInfo(dummy, "Change domain radius to %g for grid %" PetscInt_FMT "\n", (double)ctx->radius[grid], grid));
1288: }
1289: ctx->radius[grid] *= ctx->thermal_speed[grid] / ctx->v_0; // scale domain by thermal radius relative to v_0
1290: ctx->radius_perp[grid] *= ctx->thermal_speed[grid] / ctx->v_0; // scale domain by thermal radius relative to v_0
1291: ctx->radius_par[grid] *= ctx->thermal_speed[grid] / ctx->v_0; // scale domain by thermal radius relative to v_0
1292: }
1293: /* amr parameters */
1294: if (!fileflg) {
1295: nt = LANDAU_MAX_GRIDS;
1296: 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));
1297: 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);
1298: nt = LANDAU_MAX_GRIDS;
1299: PetscCall(PetscOptionsIntArray("-dm_landau_amr_post_refine", "Number of levels to uniformly refine after AMR", "plexland.c", ctx->postAMRRefine, &nt, &flg));
1300: for (ii = 1; ii < ctx->num_grids; ii++) ctx->postAMRRefine[ii] = ctx->postAMRRefine[0]; // all grids the same now
1301: 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));
1302: 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));
1303: 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));
1304: 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));
1305: 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));
1306: 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));
1307: /* spherical domain */
1308: if (ctx->sphere || ctx->simplex) {
1309: ctx->sphere_uniform_normal = PETSC_FALSE;
1310: 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));
1311: if (!ctx->sphere_uniform_normal) { // true
1312: nt = LANDAU_MAX_GRIDS;
1313: 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));
1314: if (flg && nt < ctx->num_grids) {
1315: for (PetscInt grid = nt; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_90degree[grid] = ctx->sphere_inner_radius_90degree[0];
1316: } else if (!flg || nt == 0) {
1317: if (LANDAU_DIM == 2) {
1318: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_90degree[grid] = 0.4; // optimized for R=5, Q4, AMR=0
1319: } else {
1320: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_90degree[grid] = 0.577 * 0.40;
1321: }
1322: }
1323: nt = LANDAU_MAX_GRIDS;
1324: 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));
1325: if (flg && nt < ctx->num_grids) {
1326: for (PetscInt grid = nt; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_45degree[grid] = ctx->sphere_inner_radius_45degree[0];
1327: } else if (!flg || nt == 0) {
1328: if (LANDAU_DIM == 2) {
1329: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_45degree[grid] = 0.45; // optimized for R=5, Q4, AMR=0
1330: } else {
1331: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_45degree[grid] = 0.4; // 3D sphere
1332: }
1333: }
1334: 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]));
1335: } else {
1336: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1337: switch (ctx->numAMRRefine[grid]) {
1338: case 0:
1339: case 1:
1340: case 2:
1341: case 3:
1342: default:
1343: if (LANDAU_DIM == 2) {
1344: ctx->sphere_inner_radius_90degree[grid] = 0.40;
1345: ctx->sphere_inner_radius_45degree[grid] = 0.45;
1346: } else {
1347: ctx->sphere_inner_radius_45degree[grid] = 0.25;
1348: }
1349: }
1350: }
1351: }
1352: } else {
1353: nt = LANDAU_DIM;
1354: PetscCall(PetscOptionsIntArray("-dm_landau_num_cells", "Number of cells in each dimension of base grid", "plexland.c", ctx->cells0, &nt, &flg));
1355: }
1356: }
1357: /* processing options */
1358: PetscCall(PetscOptionsBool("-dm_landau_gpu_assembly", "Assemble Jacobian on GPU", "plexland.c", ctx->gpu_assembly, &ctx->gpu_assembly, NULL));
1359: 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));
1360: 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");
1361: PetscCheck(!ctx->jacobian_field_major_order, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_jacobian_field_major_order DEPRECATED");
1362: PetscOptionsEnd();
1364: for (ii = ctx->num_species; ii < LANDAU_MAX_SPECIES; ii++) ctx->masses[ii] = ctx->thermal_temps[ii] = ctx->charges[ii] = 0;
1365: if (ctx->verbose != 0) {
1366: PetscReal pmassunit = PetscRealConstant(1.6720e-27);
1368: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "masses: e=%10.3e; ions in proton mass units: %10.3e %10.3e ...\n", (double)ctx->masses[0], (double)(ctx->masses[1] / pmassunit), (double)(ctx->num_species > 2 ? ctx->masses[2] / pmassunit : 0)));
1369: 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)));
1370: 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)));
1371: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "thermal T (K): e=%10.3e i=%10.3e %10.3e. Normalization grid %" PetscInt_FMT ": v_0=%10.3e (%10.3ec) n_0=%10.3e t_0=%10.3e %" PetscInt_FMT " batched, view batch %" PetscInt_FMT "\n", (double)ctx->thermal_temps[0],
1372: (double)ctx->thermal_temps[1], (double)((ctx->num_species > 2) ? ctx->thermal_temps[2] : 0), non_dim_grid, (double)ctx->v_0, (double)(ctx->v_0 / SPEED_OF_LIGHT), (double)ctx->n_0, (double)ctx->t_0, ctx->batch_sz, ctx->batch_view_idx));
1373: 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]));
1374: 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]));
1375: if (ctx->use_relativistic_corrections) PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\nUse relativistic corrections\n"));
1376: else PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
1377: }
1378: PetscCall(DMDestroy(&dummy));
1379: {
1380: PetscMPIInt rank;
1381: PetscCallMPI(MPI_Comm_rank(PETSC_COMM_WORLD, &rank));
1382: ctx->stage = 0;
1383: PetscCall(PetscLogEventRegister("Landau Create", DM_CLASSID, &ctx->events[13])); /* 13 */
1384: PetscCall(PetscLogEventRegister(" GPU ass. setup", DM_CLASSID, &ctx->events[2])); /* 2 */
1385: PetscCall(PetscLogEventRegister(" Build matrix", DM_CLASSID, &ctx->events[12])); /* 12 */
1386: PetscCall(PetscLogEventRegister(" Assembly maps", DM_CLASSID, &ctx->events[15])); /* 15 */
1387: PetscCall(PetscLogEventRegister("Landau Mass mat", DM_CLASSID, &ctx->events[14])); /* 14 */
1388: PetscCall(PetscLogEventRegister("Landau Operator", DM_CLASSID, &ctx->events[11])); /* 11 */
1389: PetscCall(PetscLogEventRegister("Landau Jacobian", DM_CLASSID, &ctx->events[0])); /* 0 */
1390: PetscCall(PetscLogEventRegister("Landau Mass", DM_CLASSID, &ctx->events[9])); /* 9 */
1391: PetscCall(PetscLogEventRegister(" Preamble", DM_CLASSID, &ctx->events[10])); /* 10 */
1392: PetscCall(PetscLogEventRegister(" static IP Data", DM_CLASSID, &ctx->events[7])); /* 7 */
1393: PetscCall(PetscLogEventRegister(" dynamic IP-Jac", DM_CLASSID, &ctx->events[1])); /* 1 */
1394: PetscCall(PetscLogEventRegister(" Kernel-init", DM_CLASSID, &ctx->events[3])); /* 3 */
1395: PetscCall(PetscLogEventRegister(" Jac-f-df (GPU)", DM_CLASSID, &ctx->events[8])); /* 8 */
1396: PetscCall(PetscLogEventRegister(" J Kernel (GPU)", DM_CLASSID, &ctx->events[4])); /* 4 */
1397: PetscCall(PetscLogEventRegister(" M Kernel (GPU)", DM_CLASSID, &ctx->events[16])); /* 16 */
1398: PetscCall(PetscLogEventRegister(" Copy to CPU", DM_CLASSID, &ctx->events[5])); /* 5 */
1399: PetscCall(PetscLogEventRegister(" CPU assemble", DM_CLASSID, &ctx->events[6])); /* 6 */
1401: if (rank) { /* turn off output stuff for duplicate runs - do we need to add the prefix to all this? */
1402: PetscCall(PetscOptionsClearValue(NULL, "-snes_converged_reason"));
1403: PetscCall(PetscOptionsClearValue(NULL, "-ksp_converged_reason"));
1404: PetscCall(PetscOptionsClearValue(NULL, "-snes_monitor"));
1405: PetscCall(PetscOptionsClearValue(NULL, "-ksp_monitor"));
1406: PetscCall(PetscOptionsClearValue(NULL, "-ts_monitor"));
1407: PetscCall(PetscOptionsClearValue(NULL, "-ts_view"));
1408: PetscCall(PetscOptionsClearValue(NULL, "-ts_adapt_monitor"));
1409: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_amr_dm_view"));
1410: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_amr_vec_view"));
1411: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_mass_dm_view"));
1412: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_mass_view"));
1413: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_jacobian_view"));
1414: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_mat_view"));
1415: PetscCall(PetscOptionsClearValue(NULL, "-pc_bjkokkos_ksp_converged_reason"));
1416: PetscCall(PetscOptionsClearValue(NULL, "-pc_bjkokkos_ksp_monitor"));
1417: PetscCall(PetscOptionsClearValue(NULL, "-"));
1418: PetscCall(PetscOptionsClearValue(NULL, "-info"));
1419: }
1420: }
1421: PetscFunctionReturn(PETSC_SUCCESS);
1422: }
1424: static PetscErrorCode CreateStaticData(PetscInt dim, IS grid_batch_is_inv[], LandauCtx *ctx)
1425: {
1426: PetscSection section[LANDAU_MAX_GRIDS], globsection[LANDAU_MAX_GRIDS];
1427: PetscQuadrature quad;
1428: const PetscReal *quadWeights;
1429: PetscReal invMass[LANDAU_MAX_SPECIES], nu_alpha[LANDAU_MAX_SPECIES], nu_beta[LANDAU_MAX_SPECIES];
1430: 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;
1431: PetscTabulation *Tf;
1432: PetscDS prob;
1434: PetscFunctionBegin;
1435: PetscCall(PetscFEGetDimension(ctx->fe[0], &Nb));
1436: 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);
1437: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1438: for (PetscInt ii = ctx->species_offset[grid]; ii < ctx->species_offset[grid + 1]; ii++) {
1439: invMass[ii] = ctx->m_0 / ctx->masses[ii];
1440: nu_alpha[ii] = PetscSqr(ctx->charges[ii] / ctx->m_0) * ctx->m_0 / ctx->masses[ii];
1441: nu_beta[ii] = PetscSqr(ctx->charges[ii] / ctx->epsilon0) / (8 * PETSC_PI) * ctx->t_0 * ctx->n_0 / PetscPowReal(ctx->v_0, 3);
1442: }
1443: }
1444: if (ctx->verbose == 4) {
1445: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "nu_alpha: "));
1446: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1447: PetscInt iii = ctx->species_offset[grid];
1448: for (PetscInt ii = iii; ii < ctx->species_offset[grid + 1]; ii++) PetscCall(PetscPrintf(PETSC_COMM_WORLD, " %e", (double)nu_alpha[ii]));
1449: }
1450: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\nnu_beta: "));
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_beta[ii]));
1454: }
1455: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\nnu_alpha[i]*nu_beta[j]*lambda[i][j]:\n"));
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++) {
1459: for (PetscInt gridj = 0; gridj < ctx->num_grids; gridj++) {
1460: PetscInt jjj = ctx->species_offset[gridj];
1461: 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])));
1462: }
1463: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
1464: }
1465: }
1466: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "lambda[i][j]:\n"));
1467: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1468: PetscInt iii = ctx->species_offset[grid];
1469: for (PetscInt ii = iii; ii < ctx->species_offset[grid + 1]; ii++) {
1470: for (PetscInt gridj = 0; gridj < ctx->num_grids; gridj++) {
1471: PetscInt jjj = ctx->species_offset[gridj];
1472: for (PetscInt jj = jjj; jj < ctx->species_offset[gridj + 1]; jj++) PetscCall(PetscPrintf(PETSC_COMM_WORLD, " %14.9e", (double)ctx->lambdas[grid][gridj]));
1473: }
1474: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
1475: }
1476: }
1477: }
1478: PetscCall(DMGetDS(ctx->plex[0], &prob)); // same DS for all grids
1479: PetscCall(PetscDSGetTabulation(prob, &Tf)); // Bf, &Df same for all grids
1480: /* DS, Tab and quad is same on all grids */
1481: PetscCheck(ctx->plex[0], ctx->comm, PETSC_ERR_ARG_WRONG, "Plex not created");
1482: PetscCall(PetscFEGetQuadrature(ctx->fe[0], &quad));
1483: PetscCall(PetscQuadratureGetData(quad, NULL, NULL, &Nq, NULL, &quadWeights));
1484: 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);
1485: /* setup each grid */
1486: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1487: PetscInt cStart, cEnd;
1488: PetscCheck(ctx->plex[grid] != NULL, ctx->comm, PETSC_ERR_ARG_WRONG, "Plex not created");
1489: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd));
1490: numCells[grid] = cEnd - cStart; // grids can have different topology
1491: PetscCall(DMGetLocalSection(ctx->plex[grid], §ion[grid]));
1492: PetscCall(DMGetGlobalSection(ctx->plex[grid], &globsection[grid]));
1493: PetscCall(PetscSectionGetNumFields(section[grid], &Nf[grid]));
1494: ncellsTot += numCells[grid];
1495: }
1496: /* create GPU assembly data */
1497: if (ctx->gpu_assembly) { /* we need GPU object with GPU assembly */
1498: PetscContainer container;
1499: PetscScalar *elemMatrix, *elMat;
1500: pointInterpolationP4est(*pointMaps)[LANDAU_MAX_Q_FACE];
1501: P4estVertexMaps *maps;
1502: const PetscInt *plex_batch = NULL, elMatSz = Nb * Nb * ctx->num_species * ctx->num_species;
1503: LandauIdx *coo_elem_offsets = NULL, *coo_elem_fullNb = NULL, (*coo_elem_point_offsets)[LANDAU_MAX_NQND + 1] = NULL;
1504: /* create GPU assembly data */
1505: PetscCall(PetscInfo(ctx->plex[0], "Make GPU maps %d\n", 1));
1506: PetscCall(PetscLogEventBegin(ctx->events[2], 0, 0, 0, 0));
1507: PetscCall(PetscMalloc(sizeof(*maps) * ctx->num_grids, &maps));
1508: PetscCall(PetscMalloc(sizeof(*pointMaps) * MAP_BF_SIZE, &pointMaps));
1509: PetscCall(PetscMalloc(sizeof(*elemMatrix) * elMatSz, &elemMatrix));
1511: { // setup COO assembly -- put COO metadata directly in ctx->SData_d
1512: PetscCall(PetscMalloc3(ncellsTot + 1, &coo_elem_offsets, ncellsTot, &coo_elem_fullNb, ncellsTot, &coo_elem_point_offsets)); // array of integer pointers
1513: coo_elem_offsets[0] = 0; // finish later
1514: PetscCall(PetscInfo(ctx->plex[0], "COO initialization, %" PetscInt_FMT " cells\n", ncellsTot));
1515: ctx->SData_d.coo_n_cellsTot = ncellsTot;
1516: ctx->SData_d.coo_elem_offsets = (void *)coo_elem_offsets;
1517: ctx->SData_d.coo_elem_fullNb = (void *)coo_elem_fullNb;
1518: ctx->SData_d.coo_elem_point_offsets = (void *)coo_elem_point_offsets;
1519: }
1521: ctx->SData_d.coo_max_fullnb = 0;
1522: for (PetscInt grid = 0, glb_elem_idx = 0; grid < ctx->num_grids; grid++) {
1523: PetscInt cStart, cEnd, Nfloc = Nf[grid], totDim = Nfloc * Nb;
1524: if (grid_batch_is_inv[grid]) PetscCall(ISGetIndices(grid_batch_is_inv[grid], &plex_batch));
1525: PetscCheck(!plex_batch, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_jacobian_field_major_order DEPRECATED");
1526: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd));
1527: // make maps
1528: maps[grid].d_self = NULL;
1529: maps[grid].num_elements = numCells[grid];
1530: maps[grid].num_face = (PetscInt)(pow(Nq, 1. / ((double)dim)) + .001); // Q
1531: maps[grid].num_face = (PetscInt)(pow(maps[grid].num_face, (double)(dim - 1)) + .001); // Q^2
1532: maps[grid].num_reduced = 0;
1533: maps[grid].deviceType = ctx->deviceType;
1534: maps[grid].numgrids = ctx->num_grids;
1535: // count reduced and get
1536: PetscCall(PetscMalloc(maps[grid].num_elements * sizeof(*maps[grid].gIdx), &maps[grid].gIdx));
1537: for (PetscInt ej = cStart, eidx = 0; ej < cEnd; ++ej, ++eidx, glb_elem_idx++) {
1538: if (coo_elem_offsets) coo_elem_offsets[glb_elem_idx + 1] = coo_elem_offsets[glb_elem_idx]; // start with last one, then add
1539: for (PetscInt fieldA = 0; fieldA < Nf[grid]; fieldA++) {
1540: PetscInt fullNb = 0;
1541: for (PetscInt q = 0; q < Nb; ++q) {
1542: PetscInt numindices, *indices;
1543: PetscScalar *valuesOrig = elMat = elemMatrix;
1544: PetscCall(PetscArrayzero(elMat, totDim * totDim));
1545: elMat[(fieldA * Nb + q) * totDim + fieldA * Nb + q] = 1;
1546: PetscCall(DMPlexGetClosureIndices(ctx->plex[grid], section[grid], globsection[grid], ej, PETSC_TRUE, &numindices, &indices, NULL, &elMat));
1547: if (ctx->simplex) {
1548: PetscCheck(numindices == Nb, ctx->comm, PETSC_ERR_ARG_WRONG, "numindices != Nb numindices=%" PetscInt_FMT " Nb=%" PetscInt_FMT, numindices, Nb);
1549: for (PetscInt q = 0; q < numindices; ++q) maps[grid].gIdx[eidx][fieldA][q] = indices[q];
1550: fullNb++;
1551: } else {
1552: for (PetscInt f = 0; f < numindices; ++f) { // look for a non-zero on the diagonal (is this too complicated for simplices?)
1553: if (PetscAbs(PetscRealPart(elMat[f * numindices + f])) > PETSC_MACHINE_EPSILON) {
1554: // found it
1555: if (PetscAbs(PetscRealPart(elMat[f * numindices + f] - 1.)) < PETSC_MACHINE_EPSILON) { // normal vertex 1.0
1556: if (plex_batch) {
1557: maps[grid].gIdx[eidx][fieldA][q] = plex_batch[indices[f]];
1558: } else {
1559: maps[grid].gIdx[eidx][fieldA][q] = indices[f];
1560: }
1561: fullNb++;
1562: } else { //found a constraint
1563: PetscInt jj = 0;
1564: PetscReal sum = 0;
1565: const PetscInt ff = f;
1566: maps[grid].gIdx[eidx][fieldA][q] = -maps[grid].num_reduced - 1; // store (-)index: id = -(idx+1): idx = -id - 1
1567: PetscCheck(!ctx->simplex, ctx->comm, PETSC_ERR_ARG_WRONG, "No constraints with simplex");
1568: do { // constraints are continuous in Plex - exploit that here
1569: PetscInt ii; // get 'scale'
1570: 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
1571: 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
1572: pointMaps[maps[grid].num_reduced][jj].scale += PetscRealPart(elMat[f * numindices + ff + ii]);
1573: }
1574: }
1575: sum += pointMaps[maps[grid].num_reduced][jj].scale; // diagnostic
1576: // get 'gid'
1577: if (pointMaps[maps[grid].num_reduced][jj].scale == 0) pointMaps[maps[grid].num_reduced][jj].gid = -1; // 3D has Q and Q^2 interps
1578: else {
1579: if (plex_batch) {
1580: pointMaps[maps[grid].num_reduced][jj].gid = plex_batch[indices[f]];
1581: } else {
1582: pointMaps[maps[grid].num_reduced][jj].gid = indices[f];
1583: }
1584: fullNb++;
1585: }
1586: } while (++jj < maps[grid].num_face && ++f < numindices); // jj is incremented if we hit the end
1587: while (jj < maps[grid].num_face) {
1588: pointMaps[maps[grid].num_reduced][jj].scale = 0;
1589: pointMaps[maps[grid].num_reduced][jj].gid = -1;
1590: jj++;
1591: }
1592: if (PetscAbs(sum - 1.0) > 10 * PETSC_MACHINE_EPSILON) { // debug
1593: PetscInt d, f;
1594: PetscReal tmp = 0;
1595: PetscCall(
1596: PetscPrintf(PETSC_COMM_SELF, "\t\t%" PetscInt_FMT ".%" PetscInt_FMT ".%" PetscInt_FMT ") ERROR total I = %22.16e (LANDAU_MAX_Q_FACE=%d, #face=%" PetscInt_FMT ")\n", eidx, q, fieldA, (double)sum, LANDAU_MAX_Q_FACE, maps[grid].num_face));
1597: for (d = 0, tmp = 0; d < numindices; ++d) {
1598: if (tmp != 0 && PetscAbs(tmp - 1.0) > 10 * PETSC_MACHINE_EPSILON) PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%3" PetscInt_FMT ") %3" PetscInt_FMT ": ", d, indices[d]));
1599: for (f = 0; f < numindices; ++f) tmp += PetscRealPart(elMat[d * numindices + f]);
1600: if (tmp != 0) PetscCall(PetscPrintf(ctx->comm, " | %22.16e\n", (double)tmp));
1601: }
1602: }
1603: maps[grid].num_reduced++;
1604: PetscCheck(maps[grid].num_reduced < MAP_BF_SIZE, PETSC_COMM_SELF, PETSC_ERR_PLIB, "maps[grid].num_reduced %" PetscInt_FMT " > %" PetscInt_FMT, maps[grid].num_reduced, MAP_BF_SIZE);
1605: }
1606: break;
1607: }
1608: }
1609: } // !simplex
1610: // cleanup
1611: PetscCall(DMPlexRestoreClosureIndices(ctx->plex[grid], section[grid], globsection[grid], ej, PETSC_TRUE, &numindices, &indices, NULL, &elMat));
1612: if (elMat != valuesOrig) PetscCall(DMRestoreWorkArray(ctx->plex[grid], numindices * numindices, MPIU_SCALAR, &elMat));
1613: }
1614: { // setup COO assembly
1615: coo_elem_offsets[glb_elem_idx + 1] += fullNb * fullNb; // one species block, adds a block for each species, on this element in this grid
1616: if (fieldA == 0) { // cache full Nb for this element, on this grid per species
1617: coo_elem_fullNb[glb_elem_idx] = fullNb;
1618: if (fullNb > ctx->SData_d.coo_max_fullnb) ctx->SData_d.coo_max_fullnb = fullNb;
1619: } else PetscCheck(coo_elem_fullNb[glb_elem_idx] == fullNb, PETSC_COMM_SELF, PETSC_ERR_PLIB, "full element size change with species %" PetscInt_FMT " %" PetscInt_FMT, coo_elem_fullNb[glb_elem_idx], fullNb);
1620: }
1621: } // field
1622: } // cell
1623: // allocate and copy point data maps[grid].gIdx[eidx][field][q]
1624: PetscCall(PetscMalloc(maps[grid].num_reduced * sizeof(*maps[grid].c_maps), &maps[grid].c_maps));
1625: for (PetscInt ej = 0; ej < maps[grid].num_reduced; ++ej) {
1626: for (PetscInt q = 0; q < maps[grid].num_face; ++q) {
1627: maps[grid].c_maps[ej][q].scale = pointMaps[ej][q].scale;
1628: maps[grid].c_maps[ej][q].gid = pointMaps[ej][q].gid;
1629: }
1630: }
1631: #if defined(PETSC_HAVE_KOKKOS)
1632: if (ctx->deviceType == LANDAU_KOKKOS) PetscCall(LandauKokkosCreateMatMaps(maps, pointMaps, Nf, grid)); // implies Kokkos does
1633: #endif
1634: if (plex_batch) {
1635: PetscCall(ISRestoreIndices(grid_batch_is_inv[grid], &plex_batch));
1636: PetscCall(ISDestroy(&grid_batch_is_inv[grid])); // we are done with this
1637: }
1638: } /* grids */
1639: // finish COO
1640: { // setup COO assembly
1641: PetscInt *oor, *ooc;
1642: ctx->SData_d.coo_size = coo_elem_offsets[ncellsTot] * ctx->batch_sz;
1643: PetscCall(PetscMalloc2(ctx->SData_d.coo_size, &oor, ctx->SData_d.coo_size, &ooc));
1644: for (PetscInt i = 0; i < ctx->SData_d.coo_size; i++) oor[i] = ooc[i] = -1;
1645: // get
1646: for (PetscInt grid = 0, glb_elem_idx = 0; grid < ctx->num_grids; grid++) {
1647: for (PetscInt ej = 0; ej < numCells[grid]; ++ej, glb_elem_idx++) {
1648: const PetscInt fullNb = coo_elem_fullNb[glb_elem_idx];
1649: 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
1650: coo_elem_point_offsets[glb_elem_idx][0] = 0;
1651: for (PetscInt f = 0, cnt2 = 0; f < Nb; f++) {
1652: PetscInt idx = Idxs[f];
1653: coo_elem_point_offsets[glb_elem_idx][f + 1] = coo_elem_point_offsets[glb_elem_idx][f]; // start at last
1654: if (idx >= 0) {
1655: cnt2++;
1656: coo_elem_point_offsets[glb_elem_idx][f + 1]++; // inc
1657: } else {
1658: idx = -idx - 1;
1659: for (PetscInt q = 0; q < maps[grid].num_face; q++) {
1660: if (maps[grid].c_maps[idx][q].gid < 0) break;
1661: cnt2++;
1662: coo_elem_point_offsets[glb_elem_idx][f + 1]++; // inc
1663: }
1664: }
1665: PetscCheck(cnt2 <= fullNb, PETSC_COMM_SELF, PETSC_ERR_PLIB, "wrong count %" PetscInt_FMT " < %" PetscInt_FMT, fullNb, cnt2);
1666: }
1667: PetscCheck(coo_elem_point_offsets[glb_elem_idx][Nb] == fullNb, PETSC_COMM_SELF, PETSC_ERR_PLIB, "coo_elem_point_offsets size %" PetscInt_FMT " != fullNb=%" PetscInt_FMT, coo_elem_point_offsets[glb_elem_idx][Nb], fullNb);
1668: }
1669: }
1670: // set
1671: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
1672: for (PetscInt grid = 0, glb_elem_idx = 0; grid < ctx->num_grids; grid++) {
1673: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
1674: for (PetscInt ej = 0; ej < numCells[grid]; ++ej, glb_elem_idx++) {
1675: const PetscInt fullNb = coo_elem_fullNb[glb_elem_idx], fullNb2 = fullNb * fullNb;
1676: // set (i,j)
1677: for (PetscInt fieldA = 0; fieldA < Nf[grid]; fieldA++) {
1678: const LandauIdx *const Idxs = &maps[grid].gIdx[ej][fieldA][0];
1679: PetscInt rows[LANDAU_MAX_Q_FACE], cols[LANDAU_MAX_Q_FACE];
1680: for (PetscInt f = 0; f < Nb; ++f) {
1681: const PetscInt nr = coo_elem_point_offsets[glb_elem_idx][f + 1] - coo_elem_point_offsets[glb_elem_idx][f];
1682: if (nr == 1) rows[0] = Idxs[f];
1683: else {
1684: const PetscInt idx = -Idxs[f] - 1;
1685: for (PetscInt q = 0; q < nr; q++) rows[q] = maps[grid].c_maps[idx][q].gid;
1686: }
1687: for (PetscInt g = 0; g < Nb; ++g) {
1688: const PetscInt nc = coo_elem_point_offsets[glb_elem_idx][g + 1] - coo_elem_point_offsets[glb_elem_idx][g];
1689: if (nc == 1) cols[0] = Idxs[g];
1690: else {
1691: const PetscInt idx = -Idxs[g] - 1;
1692: for (PetscInt q = 0; q < nc; q++) cols[q] = maps[grid].c_maps[idx][q].gid;
1693: }
1694: 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];
1695: for (PetscInt q = 0, idx = idx0; q < nr; q++) {
1696: for (PetscInt d = 0; d < nc; d++, idx++) {
1697: oor[idx] = rows[q] + moffset;
1698: ooc[idx] = cols[d] + moffset;
1699: }
1700: }
1701: }
1702: }
1703: }
1704: } // cell
1705: } // grid
1706: } // batch
1707: PetscCall(MatSetPreallocationCOO(ctx->J, ctx->SData_d.coo_size, oor, ooc));
1708: PetscCall(PetscFree2(oor, ooc));
1709: }
1710: PetscCall(PetscFree(pointMaps));
1711: PetscCall(PetscFree(elemMatrix));
1712: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
1713: PetscCall(PetscContainerSetPointer(container, (void *)maps));
1714: PetscCall(PetscContainerSetCtxDestroy(container, LandauGPUMapsDestroy));
1715: PetscCall(PetscObjectCompose((PetscObject)ctx->J, "assembly_maps", (PetscObject)container));
1716: PetscCall(PetscContainerDestroy(&container));
1717: PetscCall(PetscLogEventEnd(ctx->events[2], 0, 0, 0, 0));
1718: } // end GPU assembly
1719: { /* create static point data, Jacobian called first, only one vertex copy */
1720: PetscReal *invJe, *ww, *xx, *yy, *zz = NULL, *invJ_a;
1721: PetscInt outer_ipidx, outer_ej, grid, nip_glb = 0;
1722: PetscFE fe;
1723: PetscCall(PetscLogEventBegin(ctx->events[7], 0, 0, 0, 0));
1724: PetscCall(PetscInfo(ctx->plex[0], "Initialize static data\n"));
1725: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) nip_glb += Nq * numCells[grid];
1726: /* collect f data, first time is for Jacobian, but make mass now */
1727: if (ctx->verbose != 0) {
1728: PetscInt ncells = 0, N;
1729: MatInfo info;
1730: PetscCall(MatGetInfo(ctx->J, MAT_LOCAL, &info));
1731: PetscCall(MatGetSize(ctx->J, &N, NULL));
1732: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ncells += numCells[grid];
1733: 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,
1734: ctx->num_species, Nb, dim, N, (PetscInt)info.nz_used));
1735: }
1736: PetscCall(PetscMalloc4(nip_glb, &ww, nip_glb, &xx, nip_glb, &yy, nip_glb * dim * dim, &invJ_a));
1737: if (dim == 3) PetscCall(PetscMalloc1(nip_glb, &zz));
1738: if (ctx->use_energy_tensor_trick) {
1739: PetscCall(PetscFECreateDefault(PETSC_COMM_SELF, dim, 1, ctx->simplex, NULL, PETSC_DECIDE, &fe));
1740: PetscCall(PetscObjectSetName((PetscObject)fe, "energy"));
1741: }
1742: /* init each grids static data - no batch */
1743: for (grid = 0, outer_ipidx = 0, outer_ej = 0; grid < ctx->num_grids; grid++) { // OpenMP (once)
1744: Vec v2_2 = NULL; // projected function: v^2/2 for non-relativistic, gamma... for relativistic
1745: PetscSection e_section;
1746: DM dmEnergy;
1747: PetscInt cStart, cEnd, ej;
1749: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd));
1750: // prep energy trick, get v^2 / 2 vector
1751: if (ctx->use_energy_tensor_trick) {
1752: PetscErrorCode (*energyf[1])(PetscInt, PetscReal, const PetscReal[], PetscInt, PetscScalar[], void *) = {ctx->use_relativistic_corrections ? gamma_m1_f : energy_f};
1753: Vec glob_v2;
1754: PetscReal *c2_0[1], data[1] = {PetscSqr(C_0(ctx->v_0))};
1756: PetscCall(DMClone(ctx->plex[grid], &dmEnergy));
1757: PetscCall(PetscObjectSetName((PetscObject)dmEnergy, "energy"));
1758: PetscCall(DMSetField(dmEnergy, 0, NULL, (PetscObject)fe));
1759: PetscCall(DMCreateDS(dmEnergy));
1760: PetscCall(DMGetLocalSection(dmEnergy, &e_section));
1761: PetscCall(DMGetGlobalVector(dmEnergy, &glob_v2));
1762: PetscCall(PetscObjectSetName((PetscObject)glob_v2, "trick"));
1763: c2_0[0] = &data[0];
1764: PetscCall(DMProjectFunction(dmEnergy, 0., energyf, (void **)c2_0, INSERT_ALL_VALUES, glob_v2));
1765: PetscCall(DMGetLocalVector(dmEnergy, &v2_2));
1766: PetscCall(VecZeroEntries(v2_2)); /* zero BCs so don't set */
1767: PetscCall(DMGlobalToLocalBegin(dmEnergy, glob_v2, INSERT_VALUES, v2_2));
1768: PetscCall(DMGlobalToLocalEnd(dmEnergy, glob_v2, INSERT_VALUES, v2_2));
1769: PetscCall(DMViewFromOptions(dmEnergy, NULL, "-energy_dm_view"));
1770: PetscCall(VecViewFromOptions(glob_v2, NULL, "-energy_vec_view"));
1771: PetscCall(DMRestoreGlobalVector(dmEnergy, &glob_v2));
1772: }
1773: /* append part of the IP data for each grid */
1774: for (ej = 0; ej < numCells[grid]; ++ej, ++outer_ej) {
1775: PetscScalar *coefs = NULL;
1776: 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);
1777: invJe = invJ_a + outer_ej * Nq * dim * dim;
1778: PetscCall(DMPlexComputeCellGeometryFEM(ctx->plex[grid], ej + cStart, quad, vj, Jdummy, invJe, detJj));
1779: if (ctx->use_energy_tensor_trick) PetscCall(DMPlexVecGetClosure(dmEnergy, e_section, v2_2, ej + cStart, NULL, &coefs));
1780: /* create static point data */
1781: for (PetscInt qj = 0; qj < Nq; qj++, outer_ipidx++) {
1782: const PetscInt gidx = outer_ipidx;
1783: const PetscReal *invJ = &invJe[qj * dim * dim];
1784: ww[gidx] = detJj[qj] * quadWeights[qj];
1785: if (dim == 2) ww[gidx] *= vj[qj * dim + 0]; /* cylindrical coordinate, w/o 2pi */
1786: // get xx, yy, zz
1787: if (ctx->use_energy_tensor_trick) {
1788: double refSpaceDer[3], eGradPhi[3];
1789: const PetscReal *const DD = Tf[0]->T[1];
1790: const PetscReal *Dq = &DD[qj * Nb * dim];
1791: for (PetscInt d = 0; d < 3; ++d) refSpaceDer[d] = eGradPhi[d] = 0.0;
1792: for (PetscInt b = 0; b < Nb; ++b) {
1793: for (PetscInt d = 0; d < dim; ++d) refSpaceDer[d] += Dq[b * dim + d] * PetscRealPart(coefs[b]);
1794: }
1795: xx[gidx] = 1e10;
1796: if (ctx->use_relativistic_corrections) {
1797: double dg2_c2 = 0;
1798: //for (PetscInt d = 0; d < dim; ++d) refSpaceDer[d] *= c02;
1799: for (PetscInt d = 0; d < dim; ++d) dg2_c2 += PetscSqr(refSpaceDer[d]);
1800: dg2_c2 *= (double)c02;
1801: if (dg2_c2 >= .999) {
1802: xx[gidx] = vj[qj * dim + 0]; /* coordinate */
1803: yy[gidx] = vj[qj * dim + 1];
1804: if (dim == 3) zz[gidx] = vj[qj * dim + 2];
1805: 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]));
1806: } else {
1807: PetscReal fact = c02 / PetscSqrtReal(1. - dg2_c2);
1808: for (PetscInt d = 0; d < dim; ++d) refSpaceDer[d] *= fact;
1809: // could test with other point u' that (grad - grad') * U (refSpaceDer, refSpaceDer') == 0
1810: }
1811: }
1812: if (xx[gidx] == 1e10) {
1813: for (PetscInt d = 0; d < dim; ++d) {
1814: for (PetscInt e = 0; e < dim; ++e) eGradPhi[d] += invJ[e * dim + d] * refSpaceDer[e];
1815: }
1816: xx[gidx] = eGradPhi[0];
1817: yy[gidx] = eGradPhi[1];
1818: if (dim == 3) zz[gidx] = eGradPhi[2];
1819: }
1820: } else {
1821: xx[gidx] = vj[qj * dim + 0]; /* coordinate */
1822: yy[gidx] = vj[qj * dim + 1];
1823: if (dim == 3) zz[gidx] = vj[qj * dim + 2];
1824: }
1825: } /* q */
1826: if (ctx->use_energy_tensor_trick) PetscCall(DMPlexVecRestoreClosure(dmEnergy, e_section, v2_2, ej + cStart, NULL, &coefs));
1827: } /* ej */
1828: if (ctx->use_energy_tensor_trick) {
1829: PetscCall(DMRestoreLocalVector(dmEnergy, &v2_2));
1830: PetscCall(DMDestroy(&dmEnergy));
1831: }
1832: } /* grid */
1833: if (ctx->use_energy_tensor_trick) PetscCall(PetscFEDestroy(&fe));
1834: /* cache static data */
1835: if (ctx->deviceType == LANDAU_KOKKOS) {
1836: #if defined(PETSC_HAVE_KOKKOS)
1837: 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));
1838: /* free */
1839: PetscCall(PetscFree4(ww, xx, yy, invJ_a));
1840: if (dim == 3) PetscCall(PetscFree(zz));
1841: #else
1842: SETERRQ(ctx->comm, PETSC_ERR_ARG_WRONG, "-landau_device_type kokkos not built");
1843: #endif
1844: } else { /* CPU version, just copy in, only use part */
1845: 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 ?
1846: ctx->SData_d.w = (void *)ww;
1847: ctx->SData_d.x = (void *)xx;
1848: ctx->SData_d.y = (void *)yy;
1849: ctx->SData_d.z = (void *)zz;
1850: ctx->SData_d.invJ = (void *)invJ_a;
1851: 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));
1852: for (PetscInt ii = 0; ii < ctx->num_species; ii++) {
1853: nu_alpha_p[ii] = nu_alpha[ii];
1854: nu_beta_p[ii] = nu_beta[ii];
1855: invMass_p[ii] = invMass[ii];
1856: }
1857: ctx->SData_d.alpha = (void *)nu_alpha_p;
1858: ctx->SData_d.beta = (void *)nu_beta_p;
1859: ctx->SData_d.invMass = (void *)invMass_p;
1860: ctx->SData_d.lambdas = (void *)lambdas_p;
1861: for (PetscInt grid = 0; grid < LANDAU_MAX_GRIDS; grid++) {
1862: PetscReal (*lambdas)[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS] = (PetscReal (*)[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS])ctx->SData_d.lambdas;
1863: for (PetscInt gridj = 0; gridj < LANDAU_MAX_GRIDS; gridj++) (*lambdas)[grid][gridj] = ctx->lambdas[grid][gridj];
1864: }
1865: }
1866: PetscCall(PetscLogEventEnd(ctx->events[7], 0, 0, 0, 0));
1867: } // initialize
1868: PetscFunctionReturn(PETSC_SUCCESS);
1869: }
1871: /* < v, u > */
1872: 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[])
1873: {
1874: g0[0] = 1.;
1875: }
1877: /* < v, u > */
1878: 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[])
1879: {
1880: static double ttt = 1e-12;
1881: g0[0] = ttt++;
1882: }
1884: /* < v, u > */
1885: 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[])
1886: {
1887: g0[0] = 2. * PETSC_PI * x[0];
1888: }
1890: /*
1891: LandauCreateJacobianMatrix - creates ctx->J with without real data. Hard to keep sparse.
1892: - Like DMPlexLandauCreateMassMatrix. Should remove one and combine
1893: - has old support for field major ordering
1894: */
1895: static PetscErrorCode LandauCreateJacobianMatrix(MPI_Comm comm, Vec X, IS grid_batch_is_inv[LANDAU_MAX_GRIDS], LandauCtx *ctx)
1896: {
1897: PetscInt *idxs = NULL;
1898: Mat subM[LANDAU_MAX_GRIDS];
1900: PetscFunctionBegin;
1901: if (!ctx->gpu_assembly) { /* we need GPU object with GPU assembly */
1902: PetscFunctionReturn(PETSC_SUCCESS);
1903: }
1904: // get the RCM for this grid to separate out species into blocks -- create 'idxs' & 'ctx->batch_is' -- not used
1905: if (ctx->gpu_assembly && ctx->jacobian_field_major_order) PetscCall(PetscMalloc1(ctx->mat_offset[ctx->num_grids] * ctx->batch_sz, &idxs));
1906: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1907: const PetscInt *values, n = ctx->mat_offset[grid + 1] - ctx->mat_offset[grid];
1908: Mat gMat;
1909: DM massDM;
1910: PetscDS prob;
1911: Vec tvec;
1912: // get "mass" matrix for reordering
1913: PetscCall(DMClone(ctx->plex[grid], &massDM));
1914: PetscCall(DMCopyFields(ctx->plex[grid], PETSC_DETERMINE, PETSC_DETERMINE, massDM));
1915: PetscCall(DMCreateDS(massDM));
1916: PetscCall(DMGetDS(massDM, &prob));
1917: 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));
1918: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only")); // this trick is need to both sparsify the matrix and avoid runtime error
1919: PetscCall(DMCreateMatrix(massDM, &gMat));
1920: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only false"));
1921: PetscCall(MatSetOption(gMat, MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE));
1922: PetscCall(MatSetOption(gMat, MAT_IGNORE_ZERO_ENTRIES, PETSC_TRUE));
1923: PetscCall(DMCreateLocalVector(ctx->plex[grid], &tvec));
1924: PetscCall(DMPlexSNESComputeJacobianFEM(massDM, tvec, gMat, gMat, ctx));
1925: PetscCall(MatViewFromOptions(gMat, NULL, "-dm_landau_reorder_mat_view"));
1926: PetscCall(DMDestroy(&massDM));
1927: PetscCall(VecDestroy(&tvec));
1928: subM[grid] = gMat;
1929: if (ctx->gpu_assembly && ctx->jacobian_field_major_order) {
1930: MatOrderingType rtype = MATORDERINGRCM;
1931: IS isrow, isicol;
1932: PetscCall(MatGetOrdering(gMat, rtype, &isrow, &isicol));
1933: PetscCall(ISInvertPermutation(isrow, PETSC_DECIDE, &grid_batch_is_inv[grid]));
1934: PetscCall(ISGetIndices(isrow, &values));
1935: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // add batch size DMs for this species grid
1936: #if !defined(LANDAU_SPECIES_MAJOR)
1937: PetscInt N = ctx->mat_offset[ctx->num_grids], n0 = ctx->mat_offset[grid] + b_id * N;
1938: for (PetscInt ii = 0; ii < n; ++ii) idxs[n0 + ii] = values[ii] + n0;
1939: #else
1940: PetscInt n0 = ctx->mat_offset[grid] * ctx->batch_sz + b_id * n;
1941: for (PetscInt ii = 0; ii < n; ++ii) idxs[n0 + ii] = values[ii] + n0;
1942: #endif
1943: }
1944: PetscCall(ISRestoreIndices(isrow, &values));
1945: PetscCall(ISDestroy(&isrow));
1946: PetscCall(ISDestroy(&isicol));
1947: }
1948: }
1949: 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));
1950: // get a block matrix
1951: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1952: Mat B = subM[grid];
1953: PetscInt nloc, nzl, *colbuf, row, COL_BF_SIZE = 1024;
1954: PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf));
1955: PetscCall(MatGetSize(B, &nloc, NULL));
1956: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
1957: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
1958: const PetscInt *cols;
1959: const PetscScalar *vals;
1960: for (PetscInt i = 0; i < nloc; i++) {
1961: PetscCall(MatGetRow(B, i, &nzl, NULL, NULL));
1962: if (nzl > COL_BF_SIZE) {
1963: PetscCall(PetscFree(colbuf));
1964: 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));
1965: COL_BF_SIZE = nzl;
1966: PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf));
1967: }
1968: PetscCall(MatGetRow(B, i, &nzl, &cols, &vals));
1969: for (PetscInt j = 0; j < nzl; j++) colbuf[j] = cols[j] + moffset;
1970: row = i + moffset;
1971: PetscCall(MatSetValues(ctx->J, 1, &row, nzl, colbuf, vals, INSERT_VALUES));
1972: PetscCall(MatRestoreRow(B, i, &nzl, &cols, &vals));
1973: }
1974: }
1975: PetscCall(PetscFree(colbuf));
1976: }
1977: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(MatDestroy(&subM[grid]));
1978: PetscCall(MatAssemblyBegin(ctx->J, MAT_FINAL_ASSEMBLY));
1979: PetscCall(MatAssemblyEnd(ctx->J, MAT_FINAL_ASSEMBLY));
1981: // debug
1982: PetscCall(MatViewFromOptions(ctx->J, NULL, "-dm_landau_mat_view"));
1983: if (ctx->gpu_assembly && ctx->jacobian_field_major_order) {
1984: Mat mat_block_order;
1985: PetscCall(MatCreateSubMatrix(ctx->J, ctx->batch_is, ctx->batch_is, MAT_INITIAL_MATRIX, &mat_block_order)); // use MatPermute
1986: PetscCall(MatViewFromOptions(mat_block_order, NULL, "-dm_landau_mat_view"));
1987: PetscCall(MatDestroy(&mat_block_order));
1988: PetscCall(VecScatterCreate(X, ctx->batch_is, X, NULL, &ctx->plex_batch));
1989: PetscCall(VecDuplicate(X, &ctx->work_vec));
1990: }
1991: PetscFunctionReturn(PETSC_SUCCESS);
1992: }
1994: PetscErrorCode DMPlexLandauCreateMassMatrix(DM pack, Mat *Amat);
1995: /*@C
1996: DMPlexLandauCreateVelocitySpace - Create a `DMPLEX` velocity space mesh
1998: Collective
2000: Input Parameters:
2001: + comm - The MPI communicator
2002: . dim - velocity space dimension (2 for axisymmetric, 3 for full 3X + 3V solver)
2003: - prefix - prefix for options (not tested)
2005: Output Parameters:
2006: + pack - The `DM` object representing the mesh
2007: . X - A vector (user destroys)
2008: - J - Optional matrix (object destroys)
2010: Level: beginner
2012: .seealso: `DMPlexCreate()`, `DMPlexLandauDestroyVelocitySpace()`
2013: @*/
2014: PetscErrorCode DMPlexLandauCreateVelocitySpace(MPI_Comm comm, PetscInt dim, const char prefix[], Vec *X, Mat *J, DM *pack)
2015: {
2016: LandauCtx *ctx;
2017: Vec Xsub[LANDAU_MAX_GRIDS];
2018: IS grid_batch_is_inv[LANDAU_MAX_GRIDS];
2020: PetscFunctionBegin;
2021: PetscCheck(dim == 2 || dim == 3, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Only 2D and 3D supported");
2022: PetscCheck(LANDAU_DIM == dim, PETSC_COMM_SELF, PETSC_ERR_PLIB, "dim %" PetscInt_FMT " != LANDAU_DIM %d", dim, LANDAU_DIM);
2023: PetscCall(PetscNew(&ctx));
2024: ctx->comm = comm; /* used for diagnostics and global errors */
2025: /* process options */
2026: PetscCall(ProcessOptions(ctx, prefix));
2027: if (dim == 2) ctx->use_relativistic_corrections = PETSC_FALSE;
2028: /* Create Mesh */
2029: PetscCall(DMCompositeCreate(PETSC_COMM_SELF, pack));
2030: PetscCall(PetscLogEventBegin(ctx->events[13], 0, 0, 0, 0));
2031: PetscCall(PetscLogEventBegin(ctx->events[15], 0, 0, 0, 0));
2032: PetscCall(LandauDMCreateVMeshes(PETSC_COMM_SELF, dim, prefix, ctx, *pack)); // creates grids (Forest of AMR)
2033: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2034: /* create FEM */
2035: PetscCall(SetupDS(ctx->plex[grid], dim, grid, ctx));
2036: /* set initial state */
2037: PetscCall(DMCreateGlobalVector(ctx->plex[grid], &Xsub[grid]));
2038: PetscCall(PetscObjectSetName((PetscObject)Xsub[grid], "u_orig"));
2039: /* initial static refinement, no solve */
2040: PetscCall(LandauSetInitialCondition(ctx->plex[grid], Xsub[grid], grid, 0, 1, ctx));
2041: /* forest refinement - forest goes in (if forest), plex comes out */
2042: if (ctx->use_p4est) {
2043: DM plex;
2044: PetscCall(adapt(grid, ctx, &Xsub[grid])); // forest goes in, plex comes out
2045: if (grid == 0) {
2046: PetscCall(DMViewFromOptions(ctx->plex[grid], NULL, "-dm_landau_amr_dm_view")); // need to differentiate - todo
2047: PetscCall(VecViewFromOptions(Xsub[grid], NULL, "-dm_landau_amr_vec_view"));
2048: }
2049: // convert to plex, all done with this level
2050: PetscCall(DMConvert(ctx->plex[grid], DMPLEX, &plex));
2051: PetscCall(DMDestroy(&ctx->plex[grid]));
2052: ctx->plex[grid] = plex;
2053: }
2054: #if !defined(LANDAU_SPECIES_MAJOR)
2055: PetscCall(DMCompositeAddDM(*pack, ctx->plex[grid]));
2056: #else
2057: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // add batch size DMs for this species grid
2058: PetscCall(DMCompositeAddDM(*pack, ctx->plex[grid]));
2059: }
2060: #endif
2061: PetscCall(DMSetApplicationContext(ctx->plex[grid], ctx));
2062: }
2063: #if !defined(LANDAU_SPECIES_MAJOR)
2064: // stack the batched DMs, could do it all here!!! b_id=0
2065: for (PetscInt b_id = 1; b_id < ctx->batch_sz; b_id++) {
2066: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMCompositeAddDM(*pack, ctx->plex[grid]));
2067: }
2068: #endif
2069: // create ctx->mat_offset
2070: ctx->mat_offset[0] = 0;
2071: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2072: PetscInt n;
2073: PetscCall(VecGetLocalSize(Xsub[grid], &n));
2074: ctx->mat_offset[grid + 1] = ctx->mat_offset[grid] + n;
2075: }
2076: // creat DM & Jac
2077: PetscCall(DMSetApplicationContext(*pack, ctx));
2078: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only"));
2079: PetscCall(DMCreateMatrix(*pack, &ctx->J));
2080: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only false"));
2081: PetscCall(MatSetOption(ctx->J, MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE));
2082: PetscCall(MatSetOption(ctx->J, MAT_IGNORE_ZERO_ENTRIES, PETSC_TRUE));
2083: PetscCall(PetscObjectSetName((PetscObject)ctx->J, "Jac"));
2084: // construct initial conditions in X
2085: PetscCall(DMCreateGlobalVector(*pack, X));
2086: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2087: PetscInt n;
2088: PetscCall(VecGetLocalSize(Xsub[grid], &n));
2089: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
2090: PetscScalar const *values;
2091: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
2092: PetscCall(LandauSetInitialCondition(ctx->plex[grid], Xsub[grid], grid, b_id, ctx->batch_sz, ctx));
2093: PetscCall(VecGetArrayRead(Xsub[grid], &values)); // Drop whole grid in Plex ordering
2094: for (PetscInt i = 0, idx = moffset; i < n; i++, idx++) PetscCall(VecSetValue(*X, idx, values[i], INSERT_VALUES));
2095: PetscCall(VecRestoreArrayRead(Xsub[grid], &values));
2096: }
2097: }
2098: // cleanup
2099: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(VecDestroy(&Xsub[grid]));
2100: /* check for correct matrix type */
2101: if (ctx->gpu_assembly) { /* we need GPU object with GPU assembly */
2102: PetscBool flg;
2103: if (ctx->deviceType == LANDAU_KOKKOS) {
2104: PetscCall(PetscObjectTypeCompareAny((PetscObject)ctx->J, &flg, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, MATAIJKOKKOS, ""));
2105: #if defined(PETSC_HAVE_KOKKOS)
2106: 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'");
2107: #else
2108: 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'");
2109: #endif
2110: }
2111: }
2112: PetscCall(PetscLogEventEnd(ctx->events[15], 0, 0, 0, 0));
2114: // create field major ordering
2115: ctx->work_vec = NULL;
2116: ctx->plex_batch = NULL;
2117: ctx->batch_is = NULL;
2118: for (PetscInt i = 0; i < LANDAU_MAX_GRIDS; i++) grid_batch_is_inv[i] = NULL;
2119: PetscCall(PetscLogEventBegin(ctx->events[12], 0, 0, 0, 0));
2120: PetscCall(LandauCreateJacobianMatrix(comm, *X, grid_batch_is_inv, ctx));
2121: PetscCall(PetscLogEventEnd(ctx->events[12], 0, 0, 0, 0));
2123: // create AMR GPU assembly maps and static GPU data
2124: PetscCall(CreateStaticData(dim, grid_batch_is_inv, ctx));
2126: PetscCall(PetscLogEventEnd(ctx->events[13], 0, 0, 0, 0));
2128: // create mass matrix
2129: PetscCall(DMPlexLandauCreateMassMatrix(*pack, NULL));
2131: if (J) *J = ctx->J;
2133: if (ctx->gpu_assembly && ctx->jacobian_field_major_order) {
2134: PetscContainer container;
2135: // cache ctx for KSP with batch/field major Jacobian ordering -ksp_type gmres/etc -dm_landau_jacobian_field_major_order
2136: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
2137: PetscCall(PetscContainerSetPointer(container, (void *)ctx));
2138: PetscCall(PetscObjectCompose((PetscObject)ctx->J, "LandauCtx", (PetscObject)container));
2139: PetscCall(PetscContainerDestroy(&container));
2140: // batch solvers need to map -- can batch solvers work
2141: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
2142: PetscCall(PetscContainerSetPointer(container, (void *)ctx->plex_batch));
2143: PetscCall(PetscObjectCompose((PetscObject)ctx->J, "plex_batch_is", (PetscObject)container));
2144: PetscCall(PetscContainerDestroy(&container));
2145: }
2146: // for batch solvers
2147: {
2148: PetscContainer container;
2149: PetscInt *pNf;
2150: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
2151: PetscCall(PetscMalloc1(sizeof(*pNf), &pNf));
2152: *pNf = ctx->batch_sz;
2153: PetscCall(PetscContainerSetPointer(container, (void *)pNf));
2154: PetscCall(PetscContainerSetCtxDestroy(container, PetscCtxDestroyDefault));
2155: PetscCall(PetscObjectCompose((PetscObject)ctx->J, "batch size", (PetscObject)container));
2156: PetscCall(PetscContainerDestroy(&container));
2157: }
2158: PetscFunctionReturn(PETSC_SUCCESS);
2159: }
2161: /*@C
2162: DMPlexLandauAccess - Access to the distribution function with user callback
2164: Collective
2166: Input Parameters:
2167: + pack - the `DMCOMPOSITE`
2168: . func - call back function
2169: - user_ctx - user context
2171: Input/Output Parameter:
2172: . X - Vector to data to
2174: Level: advanced
2176: .seealso: `DMPlexLandauCreateVelocitySpace()`
2177: @*/
2178: PetscErrorCode DMPlexLandauAccess(DM pack, Vec X, PetscErrorCode (*func)(DM, Vec, PetscInt, PetscInt, PetscInt, void *), void *user_ctx)
2179: {
2180: LandauCtx *ctx;
2182: PetscFunctionBegin;
2183: PetscCall(DMGetApplicationContext(pack, &ctx)); // uses ctx->num_grids; ctx->plex[grid]; ctx->batch_sz; ctx->mat_offset
2184: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2185: PetscInt dim, n;
2186: PetscCall(DMGetDimension(pack, &dim));
2187: for (PetscInt sp = ctx->species_offset[grid], i0 = 0; sp < ctx->species_offset[grid + 1]; sp++, i0++) {
2188: Vec vec;
2189: PetscInt vf[1] = {i0};
2190: IS vis;
2191: DM vdm;
2192: PetscCall(DMCreateSubDM(ctx->plex[grid], 1, vf, &vis, &vdm));
2193: PetscCall(DMSetApplicationContext(vdm, ctx)); // the user might want this
2194: PetscCall(DMCreateGlobalVector(vdm, &vec));
2195: PetscCall(VecGetSize(vec, &n));
2196: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
2197: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
2198: PetscCall(VecZeroEntries(vec));
2199: /* Add your data with 'dm' for species 'sp' to 'vec' */
2200: PetscCall(func(vdm, vec, i0, grid, b_id, user_ctx));
2201: /* add to global */
2202: PetscScalar const *values;
2203: const PetscInt *offsets;
2204: PetscCall(VecGetArrayRead(vec, &values));
2205: PetscCall(ISGetIndices(vis, &offsets));
2206: for (PetscInt i = 0; i < n; i++) PetscCall(VecSetValue(X, moffset + offsets[i], values[i], ADD_VALUES));
2207: PetscCall(VecRestoreArrayRead(vec, &values));
2208: PetscCall(ISRestoreIndices(vis, &offsets));
2209: } // batch
2210: PetscCall(VecDestroy(&vec));
2211: PetscCall(ISDestroy(&vis));
2212: PetscCall(DMDestroy(&vdm));
2213: }
2214: } // grid
2215: PetscFunctionReturn(PETSC_SUCCESS);
2216: }
2218: /*@
2219: DMPlexLandauDestroyVelocitySpace - Destroy a `DMPLEX` velocity space mesh
2221: Collective
2223: Input/Output Parameters:
2224: . dm - the `DM` to destroy
2226: Level: beginner
2228: .seealso: `DMPlexLandauCreateVelocitySpace()`
2229: @*/
2230: PetscErrorCode DMPlexLandauDestroyVelocitySpace(DM *dm)
2231: {
2232: LandauCtx *ctx;
2234: PetscFunctionBegin;
2235: PetscCall(DMGetApplicationContext(*dm, &ctx));
2236: PetscCall(MatDestroy(&ctx->M));
2237: PetscCall(MatDestroy(&ctx->J));
2238: for (PetscInt ii = 0; ii < ctx->num_species; ii++) PetscCall(PetscFEDestroy(&ctx->fe[ii]));
2239: PetscCall(ISDestroy(&ctx->batch_is));
2240: PetscCall(VecDestroy(&ctx->work_vec));
2241: PetscCall(VecScatterDestroy(&ctx->plex_batch));
2242: if (ctx->deviceType == LANDAU_KOKKOS) {
2243: #if defined(PETSC_HAVE_KOKKOS)
2244: PetscCall(LandauKokkosStaticDataClear(&ctx->SData_d));
2245: #else
2246: SETERRQ(ctx->comm, PETSC_ERR_ARG_WRONG, "-landau_device_type %s not built", "kokkos");
2247: #endif
2248: } else {
2249: if (ctx->SData_d.x) { /* in a CPU run */
2250: 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;
2251: 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;
2252: PetscCall(PetscFree4(ww, xx, yy, invJ));
2253: if (zz) PetscCall(PetscFree(zz));
2254: if (coo_elem_offsets) PetscCall(PetscFree3(coo_elem_offsets, coo_elem_fullNb, coo_elem_point_offsets)); // could be NULL
2255: PetscCall(PetscFree4(ctx->SData_d.alpha, ctx->SData_d.beta, ctx->SData_d.invMass, ctx->SData_d.lambdas));
2256: }
2257: }
2259: if (ctx->times[LANDAU_MATRIX_TOTAL] > 0) { // OMP timings
2260: PetscCall(PetscPrintf(ctx->comm, "TSStep N 1.0 %10.3e\n", ctx->times[LANDAU_EX2_TSSOLVE]));
2261: 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));
2262: PetscCall(PetscPrintf(ctx->comm, "3: Landau: %10.3e\n", ctx->times[LANDAU_MATRIX_TOTAL]));
2263: PetscCall(PetscPrintf(ctx->comm, "Landau Jacobian %" PetscInt_FMT " 1.0 %10.3e\n", (PetscInt)ctx->times[LANDAU_JACOBIAN_COUNT], ctx->times[LANDAU_JACOBIAN]));
2264: PetscCall(PetscPrintf(ctx->comm, "Landau Operator N 1.0 %10.3e\n", ctx->times[LANDAU_OPERATOR]));
2265: PetscCall(PetscPrintf(ctx->comm, "Landau Mass N 1.0 %10.3e\n", ctx->times[LANDAU_MASS]));
2266: PetscCall(PetscPrintf(ctx->comm, " Jac-f-df (GPU) N 1.0 %10.3e\n", ctx->times[LANDAU_F_DF]));
2267: PetscCall(PetscPrintf(ctx->comm, " Kernel (GPU) N 1.0 %10.3e\n", ctx->times[LANDAU_KERNEL]));
2268: PetscCall(PetscPrintf(ctx->comm, "MatLUFactorNum X 1.0 %10.3e\n", ctx->times[KSP_FACTOR]));
2269: PetscCall(PetscPrintf(ctx->comm, "MatSolve X 1.0 %10.3e\n", ctx->times[KSP_SOLVE]));
2270: }
2271: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMDestroy(&ctx->plex[grid]));
2272: PetscCall(PetscFree(ctx));
2273: PetscCall(DMDestroy(dm));
2274: PetscFunctionReturn(PETSC_SUCCESS);
2275: }
2277: /* < v, ru > */
2278: 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)
2279: {
2280: PetscInt ii = (PetscInt)PetscRealPart(constants[0]);
2281: f0[0] = u[ii];
2282: }
2284: /* < v, ru > */
2285: 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)
2286: {
2287: PetscInt ii = (PetscInt)PetscRealPart(constants[0]), jj = (PetscInt)PetscRealPart(constants[1]);
2288: f0[0] = x[jj] * u[ii]; /* x momentum */
2289: }
2291: 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)
2292: {
2293: PetscInt i, ii = (PetscInt)PetscRealPart(constants[0]);
2294: double tmp1 = 0.;
2295: for (i = 0; i < dim; ++i) tmp1 += x[i] * x[i];
2296: f0[0] = tmp1 * u[ii];
2297: }
2299: static PetscErrorCode gamma_n_f(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf, PetscScalar *u, void *actx)
2300: {
2301: const PetscReal *c2_0_arr = ((PetscReal *)actx);
2302: const PetscReal c02 = c2_0_arr[0];
2304: PetscFunctionBegin;
2305: for (PetscInt s = 0; s < Nf; s++) {
2306: PetscReal tmp1 = 0.;
2307: for (PetscInt i = 0; i < dim; ++i) tmp1 += x[i] * x[i];
2308: #if defined(PETSC_USE_DEBUG)
2309: u[s] = PetscSqrtReal(1. + tmp1 / c02); // u[0] = PetscSqrtReal(1. + xx);
2310: #else
2311: {
2312: PetscReal xx = tmp1 / c02;
2313: u[s] = xx / (PetscSqrtReal(1. + xx) + 1.); // better conditioned = xx/(PetscSqrtReal(1. + xx) + 1.)
2314: }
2315: #endif
2316: }
2317: PetscFunctionReturn(PETSC_SUCCESS);
2318: }
2320: /* < v, ru > */
2321: 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)
2322: {
2323: PetscInt ii = (PetscInt)PetscRealPart(constants[0]);
2324: f0[0] = 2. * PETSC_PI * x[0] * u[ii];
2325: }
2327: /* < v, ru > */
2328: 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)
2329: {
2330: PetscInt ii = (PetscInt)PetscRealPart(constants[0]);
2331: f0[0] = 2. * PETSC_PI * x[0] * x[1] * u[ii];
2332: }
2334: 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)
2335: {
2336: PetscInt ii = (PetscInt)PetscRealPart(constants[0]);
2337: f0[0] = 2. * PETSC_PI * x[0] * (x[0] * x[0] + x[1] * x[1]) * u[ii];
2338: }
2340: /*@
2341: DMPlexLandauPrintNorms - collects moments and prints them
2343: Collective
2345: Input Parameters:
2346: + X - the state
2347: - stepi - current step to print
2349: Level: beginner
2351: .seealso: `DMPlexLandauCreateVelocitySpace()`
2352: @*/
2353: PetscErrorCode DMPlexLandauPrintNorms(Vec X, PetscInt stepi)
2354: {
2355: LandauCtx *ctx;
2356: PetscDS prob;
2357: DM pack;
2358: PetscInt cStart, cEnd, dim, ii, i0, nDMs;
2359: PetscScalar xmomentumtot = 0, ymomentumtot = 0, zmomentumtot = 0, energytot = 0, densitytot = 0, tt[LANDAU_MAX_SPECIES];
2360: PetscScalar xmomentum[LANDAU_MAX_SPECIES], ymomentum[LANDAU_MAX_SPECIES], zmomentum[LANDAU_MAX_SPECIES], energy[LANDAU_MAX_SPECIES], density[LANDAU_MAX_SPECIES];
2361: Vec *globXArray;
2363: PetscFunctionBegin;
2364: PetscCall(VecGetDM(X, &pack));
2365: PetscCheck(pack, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Vector has no DM");
2366: PetscCall(DMGetDimension(pack, &dim));
2367: PetscCheck(dim == 2 || dim == 3, PETSC_COMM_SELF, PETSC_ERR_PLIB, "dim %" PetscInt_FMT " not in [2,3]", dim);
2368: PetscCall(DMGetApplicationContext(pack, &ctx));
2369: PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context");
2370: /* print momentum and energy */
2371: PetscCall(DMCompositeGetNumberDM(pack, &nDMs));
2372: 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);
2373: PetscCall(PetscMalloc(sizeof(*globXArray) * nDMs, &globXArray));
2374: PetscCall(DMCompositeGetAccessArray(pack, X, nDMs, NULL, globXArray));
2375: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2376: Vec Xloc = globXArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)];
2377: PetscCall(DMGetDS(ctx->plex[grid], &prob));
2378: for (ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) {
2379: PetscScalar user[2] = {(PetscScalar)i0, ctx->charges[ii]};
2380: PetscCall(PetscDSSetConstants(prob, 2, user));
2381: if (dim == 2) { /* 2/3X + 3V (cylindrical coordinates) */
2382: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_rden));
2383: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2384: density[ii] = tt[0] * ctx->n_0 * ctx->charges[ii];
2385: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_rmom));
2386: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2387: zmomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii];
2388: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_rv2));
2389: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2390: energy[ii] = tt[0] * 0.5 * ctx->n_0 * ctx->v_0 * ctx->v_0 * ctx->masses[ii];
2391: zmomentumtot += zmomentum[ii];
2392: energytot += energy[ii];
2393: densitytot += density[ii];
2394: 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])));
2395: } else { /* 2/3Xloc + 3V */
2396: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_den));
2397: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2398: density[ii] = tt[0] * ctx->n_0 * ctx->charges[ii];
2399: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_mom));
2400: user[1] = 0;
2401: PetscCall(PetscDSSetConstants(prob, 2, user));
2402: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2403: xmomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii];
2404: user[1] = 1;
2405: PetscCall(PetscDSSetConstants(prob, 2, user));
2406: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2407: ymomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii];
2408: user[1] = 2;
2409: PetscCall(PetscDSSetConstants(prob, 2, user));
2410: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2411: zmomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii];
2412: if (ctx->use_relativistic_corrections) {
2413: /* gamma * M * f */
2414: if (ii == 0 && grid == 0) { // do all at once
2415: Vec Mf, globGamma, *globMfArray, *globGammaArray;
2416: PetscErrorCode (*gammaf[1])(PetscInt, PetscReal, const PetscReal[], PetscInt, PetscScalar[], void *) = {gamma_n_f};
2417: PetscReal *c2_0[1], data[1];
2419: PetscCall(VecDuplicate(X, &globGamma));
2420: PetscCall(VecDuplicate(X, &Mf));
2421: PetscCall(PetscMalloc(sizeof(*globMfArray) * nDMs, &globMfArray));
2422: PetscCall(PetscMalloc(sizeof(*globMfArray) * nDMs, &globGammaArray));
2423: /* M * f */
2424: PetscCall(MatMult(ctx->M, X, Mf));
2425: /* gamma */
2426: PetscCall(DMCompositeGetAccessArray(pack, globGamma, nDMs, NULL, globGammaArray));
2427: 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
2428: Vec v1 = globGammaArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)];
2429: data[0] = PetscSqr(C_0(ctx->v_0));
2430: c2_0[0] = &data[0];
2431: PetscCall(DMProjectFunction(ctx->plex[grid], 0., gammaf, (void **)c2_0, INSERT_ALL_VALUES, v1));
2432: }
2433: PetscCall(DMCompositeRestoreAccessArray(pack, globGamma, nDMs, NULL, globGammaArray));
2434: /* gamma * Mf */
2435: PetscCall(DMCompositeGetAccessArray(pack, globGamma, nDMs, NULL, globGammaArray));
2436: PetscCall(DMCompositeGetAccessArray(pack, Mf, nDMs, NULL, globMfArray));
2437: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { // yes a grid loop in a grid loop to print nice
2438: PetscInt Nf = ctx->species_offset[grid + 1] - ctx->species_offset[grid], N, bs;
2439: Vec Mfsub = globMfArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)], Gsub = globGammaArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)], v1, v2;
2440: // get each component
2441: PetscCall(VecGetSize(Mfsub, &N));
2442: PetscCall(VecCreate(ctx->comm, &v1));
2443: PetscCall(VecSetSizes(v1, PETSC_DECIDE, N / Nf));
2444: PetscCall(VecCreate(ctx->comm, &v2));
2445: PetscCall(VecSetSizes(v2, PETSC_DECIDE, N / Nf));
2446: PetscCall(VecSetFromOptions(v1)); // ???
2447: PetscCall(VecSetFromOptions(v2));
2448: // get each component
2449: PetscCall(VecGetBlockSize(Gsub, &bs));
2450: PetscCheck(bs == Nf, PETSC_COMM_SELF, PETSC_ERR_PLIB, "bs %" PetscInt_FMT " != num_species %" PetscInt_FMT " in Gsub", bs, Nf);
2451: PetscCall(VecGetBlockSize(Mfsub, &bs));
2452: PetscCheck(bs == Nf, PETSC_COMM_SELF, PETSC_ERR_PLIB, "bs %" PetscInt_FMT " != num_species %" PetscInt_FMT, bs, Nf);
2453: for (PetscInt i = 0, ix = ctx->species_offset[grid]; i < Nf; i++, ix++) {
2454: PetscScalar val;
2455: PetscCall(VecStrideGather(Gsub, i, v1, INSERT_VALUES)); // this is not right -- TODO
2456: PetscCall(VecStrideGather(Mfsub, i, v2, INSERT_VALUES));
2457: PetscCall(VecDot(v1, v2, &val));
2458: energy[ix] = PetscRealPart(val) * ctx->n_0 * ctx->v_0 * ctx->v_0 * ctx->masses[ix];
2459: }
2460: PetscCall(VecDestroy(&v1));
2461: PetscCall(VecDestroy(&v2));
2462: } /* grids */
2463: PetscCall(DMCompositeRestoreAccessArray(pack, globGamma, nDMs, NULL, globGammaArray));
2464: PetscCall(DMCompositeRestoreAccessArray(pack, Mf, nDMs, NULL, globMfArray));
2465: PetscCall(PetscFree(globGammaArray));
2466: PetscCall(PetscFree(globMfArray));
2467: PetscCall(VecDestroy(&globGamma));
2468: PetscCall(VecDestroy(&Mf));
2469: }
2470: } else {
2471: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_v2));
2472: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2473: energy[ii] = 0.5 * tt[0] * ctx->n_0 * ctx->v_0 * ctx->v_0 * ctx->masses[ii];
2474: }
2475: 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])));
2476: xmomentumtot += xmomentum[ii];
2477: ymomentumtot += ymomentum[ii];
2478: zmomentumtot += zmomentum[ii];
2479: energytot += energy[ii];
2480: densitytot += density[ii];
2481: }
2482: if (ctx->num_species > 1) PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
2483: }
2484: }
2485: PetscCall(DMCompositeRestoreAccessArray(pack, X, nDMs, NULL, globXArray));
2486: PetscCall(PetscFree(globXArray));
2487: /* totals */
2488: PetscCall(DMPlexGetHeightStratum(ctx->plex[0], 0, &cStart, &cEnd));
2489: if (ctx->num_species > 1) {
2490: if (dim == 2) {
2491: 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),
2492: (double)(ctx->masses[1] / ctx->masses[0]), cEnd - cStart));
2493: } else {
2494: 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),
2495: (double)(ctx->masses[1] / ctx->masses[0]), cEnd - cStart));
2496: }
2497: } else PetscCall(PetscPrintf(PETSC_COMM_WORLD, " -- %" PetscInt_FMT " cells", cEnd - cStart));
2498: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
2499: PetscFunctionReturn(PETSC_SUCCESS);
2500: }
2502: /*@
2503: DMPlexLandauCreateMassMatrix - Create mass matrix for Landau in Plex space (not field major order of Jacobian)
2504: - puts mass matrix into ctx->M
2506: Collective
2508: Input Parameter:
2509: . pack - the `DM` object. Puts matrix in Landau context M field
2511: Output Parameter:
2512: . Amat - The mass matrix (optional), mass matrix is added to the `DM` context
2514: Level: beginner
2516: .seealso: `DMPlexLandauCreateVelocitySpace()`
2517: @*/
2518: PetscErrorCode DMPlexLandauCreateMassMatrix(DM pack, Mat *Amat)
2519: {
2520: DM mass_pack, massDM[LANDAU_MAX_GRIDS];
2521: PetscDS prob;
2522: PetscInt ii, dim, N1 = 1, N2;
2523: LandauCtx *ctx;
2524: Mat packM, subM[LANDAU_MAX_GRIDS];
2526: PetscFunctionBegin;
2528: if (Amat) PetscAssertPointer(Amat, 2);
2529: PetscCall(DMGetApplicationContext(pack, &ctx));
2530: PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context");
2531: PetscCall(PetscLogEventBegin(ctx->events[14], 0, 0, 0, 0));
2532: PetscCall(DMGetDimension(pack, &dim));
2533: PetscCall(DMCompositeCreate(PetscObjectComm((PetscObject)pack), &mass_pack));
2534: /* create pack mass matrix */
2535: for (PetscInt grid = 0, ix = 0; grid < ctx->num_grids; grid++) {
2536: PetscCall(DMClone(ctx->plex[grid], &massDM[grid]));
2537: PetscCall(DMCopyFields(ctx->plex[grid], PETSC_DETERMINE, PETSC_DETERMINE, massDM[grid]));
2538: PetscCall(DMCreateDS(massDM[grid]));
2539: PetscCall(DMGetDS(massDM[grid], &prob));
2540: for (ix = 0, ii = ctx->species_offset[grid]; ii < ctx->species_offset[grid + 1]; ii++, ix++) {
2541: if (dim == 3) PetscCall(PetscDSSetJacobian(prob, ix, ix, g0_1, NULL, NULL, NULL));
2542: else PetscCall(PetscDSSetJacobian(prob, ix, ix, g0_r, NULL, NULL, NULL));
2543: }
2544: #if !defined(LANDAU_SPECIES_MAJOR)
2545: PetscCall(DMCompositeAddDM(mass_pack, massDM[grid]));
2546: #else
2547: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // add batch size DMs for this species grid
2548: PetscCall(DMCompositeAddDM(mass_pack, massDM[grid]));
2549: }
2550: #endif
2551: PetscCall(DMCreateMatrix(massDM[grid], &subM[grid]));
2552: }
2553: #if !defined(LANDAU_SPECIES_MAJOR)
2554: // stack the batched DMs
2555: for (PetscInt b_id = 1; b_id < ctx->batch_sz; b_id++) {
2556: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMCompositeAddDM(mass_pack, massDM[grid]));
2557: }
2558: #endif
2559: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only"));
2560: PetscCall(DMCreateMatrix(mass_pack, &packM));
2561: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only false"));
2562: PetscCall(MatSetOption(packM, MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE));
2563: PetscCall(MatSetOption(packM, MAT_IGNORE_ZERO_ENTRIES, PETSC_TRUE));
2564: PetscCall(DMDestroy(&mass_pack));
2565: /* make mass matrix for each block */
2566: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2567: Vec locX;
2568: DM plex = massDM[grid];
2569: PetscCall(DMGetLocalVector(plex, &locX));
2570: /* Mass matrix is independent of the input, so no need to fill locX */
2571: PetscCall(DMPlexSNESComputeJacobianFEM(plex, locX, subM[grid], subM[grid], ctx));
2572: PetscCall(DMRestoreLocalVector(plex, &locX));
2573: PetscCall(DMDestroy(&massDM[grid]));
2574: }
2575: PetscCall(MatGetSize(ctx->J, &N1, NULL));
2576: PetscCall(MatGetSize(packM, &N2, NULL));
2577: PetscCheck(N1 == N2, PetscObjectComm((PetscObject)pack), PETSC_ERR_PLIB, "Incorrect matrix sizes: |Jacobian| = %" PetscInt_FMT ", |Mass|=%" PetscInt_FMT, N1, N2);
2578: /* assemble block diagonals */
2579: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2580: Mat B = subM[grid];
2581: PetscInt nloc, nzl, *colbuf, COL_BF_SIZE = 1024, row;
2582: PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf));
2583: PetscCall(MatGetSize(B, &nloc, NULL));
2584: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
2585: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
2586: const PetscInt *cols;
2587: const PetscScalar *vals;
2588: for (PetscInt i = 0; i < nloc; i++) {
2589: PetscCall(MatGetRow(B, i, &nzl, NULL, NULL));
2590: if (nzl > COL_BF_SIZE) {
2591: PetscCall(PetscFree(colbuf));
2592: PetscCall(PetscInfo(pack, "Realloc buffer %" PetscInt_FMT " to %" PetscInt_FMT " (row size %" PetscInt_FMT ") \n", COL_BF_SIZE, 2 * COL_BF_SIZE, nzl));
2593: COL_BF_SIZE = nzl;
2594: PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf));
2595: }
2596: PetscCall(MatGetRow(B, i, &nzl, &cols, &vals));
2597: for (PetscInt j = 0; j < nzl; j++) colbuf[j] = cols[j] + moffset;
2598: row = i + moffset;
2599: PetscCall(MatSetValues(packM, 1, &row, nzl, colbuf, vals, INSERT_VALUES));
2600: PetscCall(MatRestoreRow(B, i, &nzl, &cols, &vals));
2601: }
2602: }
2603: PetscCall(PetscFree(colbuf));
2604: }
2605: // cleanup
2606: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(MatDestroy(&subM[grid]));
2607: PetscCall(MatAssemblyBegin(packM, MAT_FINAL_ASSEMBLY));
2608: PetscCall(MatAssemblyEnd(packM, MAT_FINAL_ASSEMBLY));
2609: PetscCall(PetscObjectSetName((PetscObject)packM, "mass"));
2610: PetscCall(MatViewFromOptions(packM, NULL, "-dm_landau_mass_view"));
2611: ctx->M = packM;
2612: if (Amat) *Amat = packM;
2613: PetscCall(PetscLogEventEnd(ctx->events[14], 0, 0, 0, 0));
2614: PetscFunctionReturn(PETSC_SUCCESS);
2615: }
2617: /*@
2618: DMPlexLandauIFunction - `TS` residual calculation, confusingly this computes the Jacobian w/o mass
2620: Collective
2622: Input Parameters:
2623: + ts - The time stepping context
2624: . time_dummy - current time (not used)
2625: . X - Current state
2626: . X_t - Time derivative of current state
2627: - actx - Landau context
2629: Output Parameter:
2630: . F - The residual
2632: Level: beginner
2634: .seealso: `DMPlexLandauCreateVelocitySpace()`, `DMPlexLandauIJacobian()`
2635: @*/
2636: PetscErrorCode DMPlexLandauIFunction(TS ts, PetscReal time_dummy, Vec X, Vec X_t, Vec F, void *actx)
2637: {
2638: LandauCtx *ctx = (LandauCtx *)actx;
2639: PetscInt dim;
2640: DM pack;
2641: #if defined(PETSC_HAVE_THREADSAFETY)
2642: double starttime, endtime;
2643: #endif
2644: PetscObjectState state;
2646: PetscFunctionBegin;
2647: PetscCall(TSGetDM(ts, &pack));
2648: PetscCall(DMGetApplicationContext(pack, &ctx));
2649: PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context");
2650: if (ctx->stage) PetscCall(PetscLogStagePush(ctx->stage));
2651: PetscCall(PetscLogEventBegin(ctx->events[11], 0, 0, 0, 0));
2652: PetscCall(PetscLogEventBegin(ctx->events[0], 0, 0, 0, 0));
2653: #if defined(PETSC_HAVE_THREADSAFETY)
2654: starttime = MPI_Wtime();
2655: #endif
2656: PetscCall(DMGetDimension(pack, &dim));
2657: PetscCall(PetscObjectStateGet((PetscObject)ctx->J, &state));
2658: if (state != ctx->norm_state) {
2659: PetscCall(MatZeroEntries(ctx->J));
2660: PetscCall(LandauFormJacobian_Internal(X, ctx->J, dim, 0.0, (void *)ctx));
2661: PetscCall(MatViewFromOptions(ctx->J, NULL, "-dm_landau_jacobian_view"));
2662: PetscCall(PetscObjectStateGet((PetscObject)ctx->J, &state));
2663: ctx->norm_state = state;
2664: } else {
2665: PetscCall(PetscInfo(ts, "WARNING Skip forming Jacobian, has not changed %" PetscInt64_FMT "\n", state));
2666: }
2667: /* mat vec for op */
2668: PetscCall(MatMult(ctx->J, X, F)); /* C*f */
2669: /* add time term */
2670: if (X_t) PetscCall(MatMultAdd(ctx->M, X_t, F, F));
2671: #if defined(PETSC_HAVE_THREADSAFETY)
2672: if (ctx->stage) {
2673: endtime = MPI_Wtime();
2674: ctx->times[LANDAU_OPERATOR] += (endtime - starttime);
2675: ctx->times[LANDAU_JACOBIAN] += (endtime - starttime);
2676: ctx->times[LANDAU_MATRIX_TOTAL] += (endtime - starttime);
2677: ctx->times[LANDAU_JACOBIAN_COUNT] += 1;
2678: }
2679: #endif
2680: PetscCall(PetscLogEventEnd(ctx->events[0], 0, 0, 0, 0));
2681: PetscCall(PetscLogEventEnd(ctx->events[11], 0, 0, 0, 0));
2682: if (ctx->stage) PetscCall(PetscLogStagePop());
2683: PetscFunctionReturn(PETSC_SUCCESS);
2684: }
2686: /*@
2687: DMPlexLandauIJacobian - `TS` Jacobian construction, confusingly this adds mass
2689: Collective
2691: Input Parameters:
2692: + ts - The time stepping context
2693: . time_dummy - current time (not used)
2694: . X - Current state
2695: . U_tdummy - Time derivative of current state (not used)
2696: . shift - shift for du/dt term
2697: - actx - Landau context
2699: Output Parameters:
2700: + Amat - Jacobian
2701: - Pmat - same as Amat
2703: Level: beginner
2705: .seealso: `DMPlexLandauCreateVelocitySpace()`, `DMPlexLandauIFunction()`
2706: @*/
2707: PetscErrorCode DMPlexLandauIJacobian(TS ts, PetscReal time_dummy, Vec X, Vec U_tdummy, PetscReal shift, Mat Amat, Mat Pmat, void *actx)
2708: {
2709: LandauCtx *ctx = NULL;
2710: PetscInt dim;
2711: DM pack;
2712: #if defined(PETSC_HAVE_THREADSAFETY)
2713: double starttime, endtime;
2714: #endif
2715: PetscObjectState state;
2717: PetscFunctionBegin;
2718: PetscCall(TSGetDM(ts, &pack));
2719: PetscCall(DMGetApplicationContext(pack, &ctx));
2720: PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context");
2721: PetscCheck(Amat == Pmat && Amat == ctx->J, ctx->comm, PETSC_ERR_PLIB, "Amat!=Pmat || Amat!=ctx->J");
2722: PetscCall(DMGetDimension(pack, &dim));
2723: /* get collision Jacobian into A */
2724: if (ctx->stage) PetscCall(PetscLogStagePush(ctx->stage));
2725: PetscCall(PetscLogEventBegin(ctx->events[11], 0, 0, 0, 0));
2726: PetscCall(PetscLogEventBegin(ctx->events[9], 0, 0, 0, 0));
2727: #if defined(PETSC_HAVE_THREADSAFETY)
2728: starttime = MPI_Wtime();
2729: #endif
2730: PetscCheck(shift != 0.0, ctx->comm, PETSC_ERR_PLIB, "zero shift");
2731: PetscCall(PetscObjectStateGet((PetscObject)ctx->J, &state));
2732: PetscCheck(state == ctx->norm_state, ctx->comm, PETSC_ERR_PLIB, "wrong state, %" PetscInt64_FMT " %" PetscInt64_FMT, ctx->norm_state, state);
2733: if (!ctx->use_matrix_mass) {
2734: PetscCall(LandauFormJacobian_Internal(X, ctx->J, dim, shift, (void *)ctx));
2735: } else { /* add mass */
2736: PetscCall(MatAXPY(Pmat, shift, ctx->M, SAME_NONZERO_PATTERN));
2737: }
2738: #if defined(PETSC_HAVE_THREADSAFETY)
2739: if (ctx->stage) {
2740: endtime = MPI_Wtime();
2741: ctx->times[LANDAU_OPERATOR] += (endtime - starttime);
2742: ctx->times[LANDAU_MASS] += (endtime - starttime);
2743: ctx->times[LANDAU_MATRIX_TOTAL] += (endtime - starttime);
2744: }
2745: #endif
2746: PetscCall(PetscLogEventEnd(ctx->events[9], 0, 0, 0, 0));
2747: PetscCall(PetscLogEventEnd(ctx->events[11], 0, 0, 0, 0));
2748: if (ctx->stage) PetscCall(PetscLogStagePop());
2749: PetscFunctionReturn(PETSC_SUCCESS);
2750: }