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(PetscCtxRt 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, &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; ei_r < numCells[grid_r]; ++ei_r) {
352: for (PetscInt qi = 0; qi < Nq; qi++, ipidx++) {
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;
369: temp1[0] += dudx[idx] * nu_beta[f + f_off] * invMass[f + f_off] * (*lambdas)[grid][grid_r];
370: temp1[1] += dudy[idx] * nu_beta[f + f_off] * invMass[f + f_off] * (*lambdas)[grid][grid_r];
371: #if LANDAU_DIM == 3
372: temp1[2] += dudz[idx] * nu_beta[f + f_off] * invMass[f + f_off] * (*lambdas)[grid][grid_r];
373: #endif
374: temp2 += ff[idx] * nu_beta[f + f_off] * (*lambdas)[grid][grid_r];
375: }
376: temp1[0] *= wi;
377: temp1[1] *= wi;
378: #if LANDAU_DIM == 3
379: temp1[2] *= wi;
380: #endif
381: temp2 *= wi;
382: #if LANDAU_DIM == 2
383: for (d2 = 0; d2 < 2; d2++) {
384: for (d3 = 0; d3 < 2; ++d3) {
385: /* K = U * grad(f): g2=e: i,A */
386: gg2_temp[d2] += Uk[d2][d3] * temp1[d3];
387: /* D = -U * (I \kron (fx)): g3=f: i,j,A */
388: gg3_temp[d2][d3] += Ud[d2][d3] * temp2;
389: }
390: }
391: #else
392: for (d2 = 0; d2 < 3; ++d2) {
393: for (d3 = 0; d3 < 3; ++d3) {
394: /* K = U * grad(f): g2 = e: i,A */
395: gg2_temp[d2] += U[d2][d3] * temp1[d3];
396: /* D = -U * (I \kron (fx)): g3 = f: i,j,A */
397: gg3_temp[d2][d3] += U[d2][d3] * temp2;
398: }
399: }
400: #endif
401: } // qi
402: } // ei_r
403: IPf_idx += nip_loc_r * Nfloc_r;
404: } /* grid_r - IPs */
405: 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);
406: // add alpha and put in gg2/3
407: for (PetscInt fieldA = 0, f_off = ctx->species_offset[grid]; fieldA < loc_Nf; ++fieldA) {
408: for (d2 = 0; d2 < LANDAU_DIM; d2++) {
409: gg2[fieldA][d2] = gg2_temp[d2] * nu_alpha[fieldA + f_off];
410: for (d3 = 0; d3 < LANDAU_DIM; d3++) gg3[fieldA][d2][d3] = -gg3_temp[d2][d3] * nu_alpha[fieldA + f_off] * invMass[fieldA + f_off];
411: }
412: }
413: /* add electric field term once per IP */
414: for (PetscInt fieldA = 0, f_off = ctx->species_offset[grid]; fieldA < loc_Nf; ++fieldA) gg2[fieldA][LANDAU_DIM - 1] += Eq_m[fieldA + f_off];
415: /* Jacobian transform - g2, g3 */
416: for (PetscInt fieldA = 0; fieldA < loc_Nf; ++fieldA) {
417: for (d = 0; d < dim; ++d) {
418: g2[fieldA][d] = 0.0;
419: for (d2 = 0; d2 < dim; ++d2) {
420: g2[fieldA][d] += invJj[d * dim + d2] * gg2[fieldA][d2];
421: g3[fieldA][d][d2] = 0.0;
422: for (d3 = 0; d3 < dim; ++d3) {
423: for (dp = 0; dp < dim; ++dp) g3[fieldA][d][d2] += invJj[d * dim + d3] * gg3[fieldA][d3][dp] * invJj[d2 * dim + dp];
424: }
425: g3[fieldA][d][d2] *= wj;
426: }
427: g2[fieldA][d] *= wj;
428: }
429: }
430: } else { // mass
431: PetscReal wj = ww[jpidx_glb];
432: /* Jacobian transform - g0 */
433: for (PetscInt fieldA = 0; fieldA < loc_Nf; ++fieldA) {
434: if (dim == 2) {
435: g0[fieldA] = wj * shift * 2. * PETSC_PI; // move this to below and remove g0
436: } else {
437: g0[fieldA] = wj * shift; // move this to below and remove g0
438: }
439: }
440: }
441: /* FE matrix construction */
442: {
443: PetscInt fieldA, d, f, d2, g;
444: const PetscReal *BJq = &BB[qj * Nb], *DIq = &DD[qj * Nb * dim];
445: /* assemble - on the diagonal (I,I) */
446: for (fieldA = 0; fieldA < loc_Nf; fieldA++) {
447: for (f = 0; f < Nb; f++) {
448: const PetscInt i = fieldA * Nb + f; /* Element matrix row */
449: for (g = 0; g < Nb; ++g) {
450: const PetscInt j = fieldA * Nb + g; /* Element matrix column */
451: const PetscInt fOff = i * totDim + j;
452: if (shift == 0.0) {
453: for (d = 0; d < dim; ++d) {
454: elemMat[fOff] += DIq[f * dim + d] * g2[fieldA][d] * BJq[g];
455: for (d2 = 0; d2 < dim; ++d2) elemMat[fOff] += DIq[f * dim + d] * g3[fieldA][d][d2] * DIq[g * dim + d2];
456: }
457: } else { // mass
458: elemMat[fOff] += BJq[f] * g0[fieldA] * BJq[g];
459: }
460: }
461: }
462: }
463: }
464: } /* qj loop */
465: if (shift == 0.0) { // Jacobian
466: PetscCall(PetscLogEventEnd(ctx->events[4], 0, 0, 0, 0));
467: } else {
468: PetscCall(PetscLogEventEnd(ctx->events[16], 0, 0, 0, 0));
469: }
470: #if defined(PETSC_HAVE_THREADSAFETY)
471: endtime = MPI_Wtime();
472: if (ctx->stage) ctx->times[LANDAU_KERNEL] += (endtime - starttime);
473: #endif
474: /* assemble matrix */
475: if (!container) {
476: PetscInt cStart;
477: PetscCall(PetscLogEventBegin(ctx->events[6], 0, 0, 0, 0));
478: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, NULL));
479: PetscCall(DMPlexMatSetClosure(ctx->plex[grid], section[grid], globsection[grid], subJ[LAND_PACK_IDX(b_id, grid)], loc_elem + cStart, elemMat, ADD_VALUES));
480: PetscCall(PetscLogEventEnd(ctx->events[6], 0, 0, 0, 0));
481: } else { // GPU like assembly for debugging
482: 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];
483: 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};
484: 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;
485: /* assemble - from the diagonal (I,I) in this format for DMPlexMatSetClosure */
486: for (fieldA = 0; fieldA < loc_Nf; fieldA++) {
487: LandauIdx *const Idxs = &maps[grid].gIdx[loc_elem][fieldA][0];
488: for (f = 0; f < Nb; f++) {
489: PetscInt idx = Idxs[f];
490: if (idx >= 0) {
491: nr = 1;
492: rows0[0] = idx;
493: row_scale[0] = 1.;
494: } else {
495: idx = -idx - 1;
496: for (q = 0, nr = 0; q < maps[grid].num_face; q++, nr++) {
497: if (maps[grid].c_maps[idx][q].gid < 0) break;
498: rows0[q] = maps[grid].c_maps[idx][q].gid;
499: row_scale[q] = maps[grid].c_maps[idx][q].scale;
500: }
501: }
502: for (g = 0; g < Nb; ++g) {
503: idx = Idxs[g];
504: if (idx >= 0) {
505: nc = 1;
506: cols0[0] = idx;
507: col_scale[0] = 1.;
508: } else {
509: idx = -idx - 1;
510: nc = maps[grid].num_face;
511: for (q = 0, nc = 0; q < maps[grid].num_face; q++, nc++) {
512: if (maps[grid].c_maps[idx][q].gid < 0) break;
513: cols0[q] = maps[grid].c_maps[idx][q].gid;
514: col_scale[q] = maps[grid].c_maps[idx][q].scale;
515: }
516: }
517: const PetscInt i = fieldA * Nb + f; /* Element matrix row */
518: const PetscInt j = fieldA * Nb + g; /* Element matrix column */
519: const PetscScalar Aij = elemMat[i * totDim + j];
520: if (coo_vals) { // mirror (i,j) in CreateStaticGPUData
521: const PetscInt fullNb = coo_elem_fullNb[glb_elem_idx], fullNb2 = fullNb * fullNb;
522: 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];
523: for (PetscInt q = 0, idx2 = idx0; q < nr; q++) {
524: for (PetscInt d = 0; d < nc; d++, idx2++) coo_vals[idx2] = row_scale[q] * col_scale[d] * Aij;
525: }
526: } else {
527: for (q = 0; q < nr; q++) rows[q] = rows0[q] + moffset;
528: for (d = 0; d < nc; d++) cols[d] = cols0[d] + moffset;
529: for (q = 0; q < nr; q++) {
530: for (d = 0; d < nc; d++) vals[q * nc + d] = row_scale[q] * col_scale[d] * Aij;
531: }
532: PetscCall(MatSetValues(JacP, nr, rows, nc, cols, vals, ADD_VALUES));
533: }
534: }
535: }
536: }
537: }
538: if (loc_elem == -1) {
539: PetscCall(PetscPrintf(ctx->comm, "CPU Element matrix\n"));
540: for (PetscInt d = 0; d < totDim; ++d) {
541: for (PetscInt f = 0; f < totDim; ++f) PetscCall(PetscPrintf(ctx->comm, " %12.5e", (double)PetscRealPart(elemMat[d * totDim + f])));
542: PetscCall(PetscPrintf(ctx->comm, "\n"));
543: }
544: exit(12);
545: }
546: PetscCall(PetscFree(elemMat));
547: } /* grid */
548: } /* outer element & batch loop */
549: if (shift == 0.0) { // mass
550: PetscCall(PetscFree4(ff, dudx, dudy, dudz));
551: }
552: if (!container) { // 'CPU' assembly move nest matrix to global JacP
553: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // OpenMP
554: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
555: 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];
556: PetscInt nloc, nzl, colbuf[1024], row;
557: const PetscInt *cols;
558: const PetscScalar *vals;
559: Mat B = subJ[LAND_PACK_IDX(b_id, grid)];
560: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
561: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
562: PetscCall(MatGetSize(B, &nloc, NULL));
563: for (PetscInt i = 0; i < nloc; i++) {
564: PetscCall(MatGetRow(B, i, &nzl, &cols, &vals));
565: PetscCheck(nzl <= 1024, PetscObjectComm((PetscObject)B), PETSC_ERR_PLIB, "Row too big: %" PetscInt_FMT, nzl);
566: for (PetscInt j = 0; j < nzl; j++) colbuf[j] = moffset + cols[j];
567: row = moffset + i;
568: PetscCall(MatSetValues(JacP, 1, &row, nzl, colbuf, vals, ADD_VALUES));
569: PetscCall(MatRestoreRow(B, i, &nzl, &cols, &vals));
570: }
571: PetscCall(MatDestroy(&B));
572: }
573: }
574: }
575: if (coo_vals) {
576: PetscCall(MatSetValuesCOO(JacP, coo_vals, ADD_VALUES));
577: PetscCall(PetscFree(coo_vals));
578: }
579: } /* CPU version */
580: PetscCall(MatAssemblyBegin(JacP, MAT_FINAL_ASSEMBLY));
581: PetscCall(MatAssemblyEnd(JacP, MAT_FINAL_ASSEMBLY));
582: /* clean up */
583: if (cellClosure) PetscCall(PetscFree(cellClosure));
584: if (xdata) PetscCall(VecRestoreArrayReadAndMemType(a_X, &xdata));
585: PetscFunctionReturn(PETSC_SUCCESS);
586: }
588: /* create DMComposite of meshes for each species group */
589: static PetscErrorCode LandauDMCreateVMeshes(MPI_Comm comm_self, const PetscInt dim, const char prefix[], LandauCtx *ctx, DM pack)
590: {
591: PetscFunctionBegin;
592: /* p4est, quads */
593: /* Create plex mesh of Landau domain */
594: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
595: PetscReal par_radius = ctx->radius_par[grid], perp_radius = ctx->radius_perp[grid];
596: if (!ctx->sphere && !ctx->simplex) { // 2 or 3D (only 3D option)
597: PetscReal lo[] = {-perp_radius, -par_radius, -par_radius}, hi[] = {perp_radius, par_radius, par_radius};
598: DMBoundaryType periodicity[3] = {DM_BOUNDARY_NONE, dim == 2 ? DM_BOUNDARY_NONE : DM_BOUNDARY_NONE, DM_BOUNDARY_NONE};
599: if (dim == 2) lo[0] = 0;
600: else {
601: lo[1] = -perp_radius;
602: hi[1] = perp_radius; // 3D y is a perp
603: }
604: 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
605: PetscCall(DMLocalizeCoordinates(ctx->plex[grid])); /* needed for periodic */
606: if (dim == 3) PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "cube"));
607: else PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "half-plane"));
608: } else if (dim == 2) {
609: size_t len;
610: PetscCall(PetscStrlen(ctx->filename, &len));
611: if (len) {
612: Vec coords;
613: PetscScalar *x;
614: PetscInt N;
615: char str[] = "-dm_landau_view_file_0";
616: str[21] += grid;
617: PetscCall(DMPlexCreateFromFile(comm_self, ctx->filename, "plexland.c", PETSC_TRUE, &ctx->plex[grid]));
618: PetscCall(DMPlexOrient(ctx->plex[grid]));
619: PetscCall(DMGetCoordinatesLocal(ctx->plex[grid], &coords));
620: PetscCall(VecGetSize(coords, &N));
621: PetscCall(VecGetArray(coords, &x));
622: /* scale by domain size */
623: for (PetscInt i = 0; i < N; i += 2) {
624: x[i + 0] *= ctx->radius_perp[grid];
625: x[i + 1] *= ctx->radius_par[grid];
626: }
627: PetscCall(VecRestoreArray(coords, &x));
628: PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], ctx->filename));
629: PetscCall(PetscInfo(ctx->plex[grid], "%" PetscInt_FMT ") Read %s mesh file (%s)\n", grid, ctx->filename, str));
630: PetscCall(DMViewFromOptions(ctx->plex[grid], NULL, str));
631: } else { // simplex forces a sphere
632: PetscInt numCells = ctx->simplex ? 12 : 6, cell_size = ctx->simplex ? 3 : 4, j;
633: const PetscInt numVerts = 11;
634: PetscInt cellsT[][4] = {
635: {0, 1, 6, 5 },
636: {1, 2, 7, 6 },
637: {2, 3, 8, 7 },
638: {3, 4, 9, 8 },
639: {5, 6, 7, 10},
640: {10, 7, 8, 9 }
641: };
642: PetscInt cellsS[][3] = {
643: {0, 1, 6 },
644: {1, 2, 6 },
645: {6, 2, 7 },
646: {7, 2, 8 },
647: {8, 2, 3 },
648: {8, 3, 4 },
649: {0, 6, 5 },
650: {5, 6, 7 },
651: {5, 7, 10},
652: {10, 7, 9 },
653: {9, 7, 8 },
654: {9, 8, 4 }
655: };
656: const PetscInt *pcell = (const PetscInt *)(ctx->simplex ? &cellsS[0][0] : &cellsT[0][0]);
657: PetscReal coords[11][2], *flatCoords = &coords[0][0];
658: PetscReal rad = ctx->radius[grid];
659: for (j = 0; j < 5; j++) { // outside edge
660: PetscReal z, r, theta = -PETSC_PI / 2 + (j % 5) * PETSC_PI / 4;
661: r = rad * PetscCosReal(theta);
662: coords[j][0] = r;
663: z = rad * PetscSinReal(theta);
664: coords[j][1] = z;
665: }
666: coords[j][0] = 0;
667: coords[j++][1] = -rad * ctx->sphere_inner_radius_90degree[grid];
668: coords[j][0] = rad * ctx->sphere_inner_radius_45degree[grid] * 0.707106781186548;
669: coords[j++][1] = -rad * ctx->sphere_inner_radius_45degree[grid] * 0.707106781186548;
670: coords[j][0] = rad * ctx->sphere_inner_radius_90degree[grid];
671: coords[j++][1] = 0;
672: coords[j][0] = rad * ctx->sphere_inner_radius_45degree[grid] * 0.707106781186548;
673: coords[j++][1] = rad * ctx->sphere_inner_radius_45degree[grid] * 0.707106781186548;
674: coords[j][0] = 0;
675: coords[j++][1] = rad * ctx->sphere_inner_radius_90degree[grid];
676: coords[j][0] = 0;
677: coords[j++][1] = 0;
678: PetscCall(DMPlexCreateFromCellListPetsc(comm_self, 2, numCells, numVerts, cell_size, ctx->interpolate, pcell, 2, flatCoords, &ctx->plex[grid]));
679: PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "semi-circle"));
680: PetscCall(PetscInfo(ctx->plex[grid], "\t%" PetscInt_FMT ") Make circle %s mesh\n", grid, ctx->simplex ? "simplex" : "tensor"));
681: }
682: } else {
683: PetscCheck(dim == 3 && ctx->sphere && !ctx->simplex, ctx->comm, PETSC_ERR_ARG_WRONG, "not: dim == 3 && ctx->sphere && !ctx->simplex");
684: PetscReal rad = ctx->radius[grid], inner_rad = rad * ctx->sphere_inner_radius_90degree[grid], outer_rad = rad;
685: const PetscInt numCells = 7, cell_size = 8, numVerts = 16;
686: const PetscInt cells[][8] = {
687: {0, 3, 2, 1, 4, 5, 6, 7 },
688: {0, 4, 5, 1, 8, 9, 13, 12},
689: {1, 5, 6, 2, 9, 10, 14, 13},
690: {2, 6, 7, 3, 10, 11, 15, 14},
691: {0, 3, 7, 4, 8, 12, 15, 11},
692: {0, 1, 2, 3, 8, 11, 10, 9 },
693: {4, 7, 6, 5, 12, 13, 14, 15}
694: };
695: PetscReal coords[16 /* numVerts */][3];
696: for (PetscInt j = 0; j < 4; j++) { // inner edge, low
697: coords[j][0] = inner_rad * (j == 0 || j == 3 ? 1 : -1);
698: coords[j][1] = inner_rad * (j / 2 < 1 ? 1 : -1);
699: coords[j][2] = inner_rad * -1;
700: }
701: for (PetscInt j = 0, jj = 4; j < 4; j++, jj++) { // inner edge, hi
702: coords[jj][0] = inner_rad * (j == 0 || j == 3 ? 1 : -1);
703: coords[jj][1] = inner_rad * (j / 2 < 1 ? 1 : -1);
704: coords[jj][2] = inner_rad * 1;
705: }
706: for (PetscInt j = 0, jj = 8; j < 4; j++, jj++) { // outer edge, low
707: coords[jj][0] = outer_rad * (j == 0 || j == 3 ? 1 : -1);
708: coords[jj][1] = outer_rad * (j / 2 < 1 ? 1 : -1);
709: coords[jj][2] = outer_rad * -1;
710: }
711: for (PetscInt j = 0, jj = 12; j < 4; j++, jj++) { // outer edge, hi
712: coords[jj][0] = outer_rad * (j == 0 || j == 3 ? 1 : -1);
713: coords[jj][1] = outer_rad * (j / 2 < 1 ? 1 : -1);
714: coords[jj][2] = outer_rad * 1;
715: }
716: PetscCall(DMPlexCreateFromCellListPetsc(comm_self, 3, numCells, numVerts, cell_size, ctx->interpolate, (const PetscInt *)cells, 3, (const PetscReal *)coords, &ctx->plex[grid]));
717: PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "cubed sphere"));
718: PetscCall(PetscInfo(ctx->plex[grid], "\t%" PetscInt_FMT ") Make cubed sphere %s mesh\n", grid, ctx->simplex ? "simplex" : "tensor"));
719: }
720: PetscCall(DMSetOptionsPrefix(ctx->plex[grid], prefix));
721: PetscCall(DMSetFromOptions(ctx->plex[grid]));
722: } // grid loop
723: PetscCall(DMSetOptionsPrefix(pack, prefix));
724: { /* convert to p4est (or whatever), wait for discretization to create pack */
725: char convType[256];
726: PetscBool flg;
728: PetscOptionsBegin(ctx->comm, prefix, "Mesh conversion options", "DMPLEX");
729: PetscCall(PetscOptionsFList("-dm_landau_type", "Convert DMPlex to another format (p4est)", "plexland.c", DMList, DMPLEX, convType, 256, &flg));
730: PetscOptionsEnd();
731: if (flg) {
732: ctx->use_p4est = PETSC_TRUE; /* flag for Forest */
733: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
734: DM dmforest;
735: PetscBool isForest;
737: PetscCall(DMConvert(ctx->plex[grid], convType, &dmforest));
738: PetscCheck(dmforest, ctx->comm, PETSC_ERR_PLIB, "Convert failed?");
739: PetscCall(DMSetOptionsPrefix(dmforest, prefix));
740: PetscCall(DMIsForest(dmforest, &isForest));
741: PetscCheck(isForest, ctx->comm, PETSC_ERR_PLIB, "Converted to non Forest?");
742: PetscCall(DMDestroy(&ctx->plex[grid]));
743: ctx->plex[grid] = dmforest; // Forest for adaptivity
744: }
745: } else ctx->use_p4est = PETSC_FALSE; /* flag for Forest */
746: }
747: PetscCall(DMSetDimension(pack, dim));
748: PetscCall(PetscObjectSetName((PetscObject)pack, "Mesh"));
749: PetscCall(DMSetApplicationContext(pack, ctx));
750: PetscFunctionReturn(PETSC_SUCCESS);
751: }
753: static PetscErrorCode SetupDS(DM pack, PetscInt dim, PetscInt grid, const char prefix[], LandauCtx *ctx)
754: {
755: PetscInt ii, i0;
756: char buf[256];
757: PetscSection section;
759: PetscFunctionBegin;
760: for (ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) {
761: if (ii == 0) PetscCall(PetscSNPrintf(buf, sizeof(buf), "e"));
762: else PetscCall(PetscSNPrintf(buf, sizeof(buf), "i%" PetscInt_FMT, ii));
763: /* Setup Discretization - FEM */
764: PetscCall(PetscFECreateDefault(PETSC_COMM_SELF, dim, 1, ctx->simplex, prefix, PETSC_DECIDE, &ctx->fe[ii]));
765: PetscCall(PetscObjectSetName((PetscObject)ctx->fe[ii], buf));
766: PetscCall(DMSetField(ctx->plex[grid], i0, NULL, (PetscObject)ctx->fe[ii]));
767: }
768: PetscCall(DMCreateDS(ctx->plex[grid]));
769: PetscCall(DMGetLocalSection(ctx->plex[grid], §ion));
770: for (PetscInt ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) {
771: if (ii == 0) PetscCall(PetscSNPrintf(buf, sizeof(buf), "se"));
772: else PetscCall(PetscSNPrintf(buf, sizeof(buf), "si%" PetscInt_FMT, ii));
773: PetscCall(PetscSectionSetComponentName(section, i0, 0, buf));
774: }
775: PetscFunctionReturn(PETSC_SUCCESS);
776: }
778: /* Define a Maxwellian function for testing out the operator. */
780: /* Using cartesian velocity space coordinates, the particle */
781: /* density, [1/m^3], is defined according to */
783: /* $$ n=\int_{R^3} dv^3 \left(\frac{m}{2\pi T}\right)^{3/2}\exp [- mv^2/(2T)] $$ */
785: /* Using some constant, c, we normalize the velocity vector into a */
786: /* dimensionless variable according to v=c*x. Thus the density, $n$, becomes */
788: /* $$ n=\int_{R^3} dx^3 \left(\frac{mc^2}{2\pi T}\right)^{3/2}\exp [- mc^2/(2T)*x^2] $$ */
790: /* Defining $\theta=2T/mc^2$, we thus find that the probability density */
791: /* for finding the particle within the interval in a box dx^3 around x is */
793: /* f(x;\theta)=\left(\frac{1}{\pi\theta}\right)^{3/2} \exp [ -x^2/\theta ] */
795: typedef struct {
796: PetscReal v_0;
797: PetscReal kT_m;
798: PetscReal n;
799: PetscReal shift;
800: } MaxwellianCtx;
802: static PetscErrorCode maxwellian(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf_dummy, PetscScalar *u, void *actx)
803: {
804: MaxwellianCtx *mctx = (MaxwellianCtx *)actx;
805: PetscInt i;
806: PetscReal v2 = 0, theta = 2 * mctx->kT_m / (mctx->v_0 * mctx->v_0), shift; /* theta = 2kT/mc^2 */
808: PetscFunctionBegin;
809: /* compute the exponents, v^2 */
810: for (i = 0; i < dim; ++i) v2 += x[i] * x[i];
811: /* evaluate the Maxwellian */
812: if (mctx->shift < 0) shift = -mctx->shift;
813: else {
814: u[0] = mctx->n * PetscPowReal(PETSC_PI * theta, -1.5) * (PetscExpReal(-v2 / theta));
815: shift = mctx->shift;
816: }
817: if (shift != 0.) {
818: v2 = 0;
819: for (i = 0; i < dim - 1; ++i) v2 += x[i] * x[i];
820: v2 += (x[dim - 1] - shift) * (x[dim - 1] - shift);
821: /* evaluate the shifted Maxwellian */
822: u[0] += mctx->n * PetscPowReal(PETSC_PI * theta, -1.5) * (PetscExpReal(-v2 / theta));
823: }
824: PetscFunctionReturn(PETSC_SUCCESS);
825: }
827: /*@
828: DMPlexLandauAddMaxwellians - Add a Maxwellian distribution to a state
830: Collective
832: Input Parameters:
833: + dm - The mesh (local)
834: . time - Current time
835: . temps - Temperatures of each species (global)
836: . ns - Number density of each species (global)
837: . grid - index into current grid - just used for offset into `temp` and `ns`
838: . b_id - batch index
839: . n_batch - number of batches
840: - actx - Landau context
842: Output Parameter:
843: . X - The state (local to this grid)
845: Level: beginner
847: .seealso: `DMPlexLandauCreateVelocitySpace()`
848: @*/
849: PetscErrorCode DMPlexLandauAddMaxwellians(DM dm, Vec X, PetscReal time, PetscReal temps[], PetscReal ns[], PetscInt grid, PetscInt b_id, PetscInt n_batch, void *actx)
850: {
851: LandauCtx *ctx = (LandauCtx *)actx;
852: PetscErrorCode (*initu[LANDAU_MAX_SPECIES])(PetscInt, PetscReal, const PetscReal[], PetscInt, PetscScalar[], void *);
853: PetscInt dim;
854: MaxwellianCtx *mctxs[LANDAU_MAX_SPECIES], data[LANDAU_MAX_SPECIES];
856: PetscFunctionBegin;
857: PetscCall(DMGetDimension(dm, &dim));
858: if (!ctx) PetscCall(DMGetApplicationContext(dm, &ctx));
859: for (PetscInt ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) {
860: mctxs[i0] = &data[i0];
861: data[i0].v_0 = ctx->v_0; // v_0 same for all grids
862: data[i0].kT_m = ctx->k * temps[ii] / ctx->masses[ii]; /* kT/m */
863: data[i0].n = ns[ii];
864: initu[i0] = maxwellian;
865: data[i0].shift = 0;
866: }
867: data[0].shift = ctx->electronShift;
868: /* need to make ADD_ALL_VALUES work - TODO */
869: PetscCall(DMProjectFunction(dm, time, initu, (void **)mctxs, INSERT_ALL_VALUES, X));
870: PetscFunctionReturn(PETSC_SUCCESS);
871: }
873: /*
874: LandauSetInitialCondition - Adds Maxwellians with context
876: Collective
878: Input Parameters:
879: . dm - The mesh
880: - grid - index into current grid - just used for offset into temp and ns
881: . b_id - batch index
882: - n_batch - number of batches
883: + actx - Landau context with T and n
885: Output Parameter:
886: . X - The state
888: Level: beginner
890: .seealso: `DMPlexLandauCreateVelocitySpace()`, `DMPlexLandauAddMaxwellians()`
891: */
892: static PetscErrorCode LandauSetInitialCondition(DM dm, Vec X, PetscInt grid, PetscInt b_id, PetscInt n_batch, void *actx)
893: {
894: LandauCtx *ctx = (LandauCtx *)actx;
896: PetscFunctionBegin;
897: if (!ctx) PetscCall(DMGetApplicationContext(dm, &ctx));
898: PetscCall(VecZeroEntries(X));
899: PetscCall(DMPlexLandauAddMaxwellians(dm, X, 0.0, ctx->thermal_temps, ctx->n, grid, b_id, n_batch, ctx));
900: PetscFunctionReturn(PETSC_SUCCESS);
901: }
903: // adapt a level once. Forest in/out
904: #if defined(PETSC_USE_INFO)
905: static const char *s_refine_names[] = {"RE", "Z1", "Origin", "Z2", "Uniform"};
906: #endif
907: static PetscErrorCode adaptToleranceFEM(PetscFE fem, Vec sol, PetscInt type, PetscInt grid, LandauCtx *ctx, DM *newForest)
908: {
909: DM forest, plex, adaptedDM = NULL;
910: PetscDS prob;
911: PetscBool isForest;
912: PetscQuadrature quad;
913: PetscInt Nq, Nb, *Nb2, cStart, cEnd, c, dim, qj, k;
914: DMLabel adaptLabel = NULL;
916: PetscFunctionBegin;
917: forest = ctx->plex[grid];
918: PetscCall(DMCreateDS(forest));
919: PetscCall(DMGetDS(forest, &prob));
920: PetscCall(DMGetDimension(forest, &dim));
921: PetscCall(DMIsForest(forest, &isForest));
922: PetscCheck(isForest, ctx->comm, PETSC_ERR_ARG_WRONG, "! Forest");
923: PetscCall(DMConvert(forest, DMPLEX, &plex));
924: PetscCall(DMPlexGetHeightStratum(plex, 0, &cStart, &cEnd));
925: PetscCall(DMLabelCreate(PETSC_COMM_SELF, "adapt", &adaptLabel));
926: PetscCall(PetscFEGetQuadrature(fem, &quad));
927: PetscCall(PetscQuadratureGetData(quad, NULL, NULL, &Nq, NULL, NULL));
928: 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);
929: PetscCall(PetscFEGetDimension(ctx->fe[0], &Nb));
930: PetscCall(PetscDSGetDimensions(prob, &Nb2));
931: PetscCheck(Nb2[0] == Nb, ctx->comm, PETSC_ERR_ARG_WRONG, " Nb = %" PetscInt_FMT " != Nb (%" PetscInt_FMT ")", Nb, Nb2[0]);
932: 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);
933: PetscCall(PetscInfo(sol, "%" PetscInt_FMT ") Refine phase: %s\n", grid, s_refine_names[type]));
934: if (type == 4) {
935: for (c = cStart; c < cEnd; c++) PetscCall(DMLabelSetValue(adaptLabel, c, DM_ADAPT_REFINE));
936: } else if (type == 2) {
937: PetscInt rCellIdx[8], nr = 0, nrmax = (dim == 3) ? 8 : 2;
938: PetscReal minRad = PETSC_INFINITY, r;
939: for (c = cStart; c < cEnd; c++) {
940: PetscReal tt, v0[LANDAU_MAX_NQND * 3], J[LANDAU_MAX_NQND * 9], invJ[LANDAU_MAX_NQND * 9], detJ[LANDAU_MAX_NQND];
941: PetscCall(DMPlexComputeCellGeometryFEM(plex, c, quad, v0, J, invJ, detJ));
942: (void)J;
943: (void)invJ;
944: for (qj = 0; qj < Nq; ++qj) {
945: tt = PetscSqr(v0[dim * qj + 0]) + PetscSqr(v0[dim * qj + 1]) + PetscSqr((dim == 3) ? v0[dim * qj + 2] : 0);
946: r = PetscSqrtReal(tt);
947: if (r < minRad - PETSC_SQRT_MACHINE_EPSILON * 10.) {
948: minRad = r;
949: nr = 0;
950: rCellIdx[nr++] = c;
951: 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));
952: } else if ((r - minRad) < PETSC_SQRT_MACHINE_EPSILON * 100. && nr < nrmax) {
953: for (k = 0; k < nr; k++)
954: if (c == rCellIdx[k]) break;
955: if (k == nr) {
956: rCellIdx[nr++] = c;
957: 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)));
958: }
959: }
960: }
961: }
962: for (k = 0; k < nr; k++) PetscCall(DMLabelSetValue(adaptLabel, rCellIdx[k], DM_ADAPT_REFINE));
963: 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));
964: } else if (type == 0 || type == 1 || type == 3) { /* refine along r=0 axis */
965: PetscScalar *coef = NULL;
966: Vec coords;
967: PetscInt csize, Nv, d, nz, nrefined = 0;
968: DM cdm;
969: PetscSection cs;
970: PetscCall(DMGetCoordinatesLocal(forest, &coords));
971: PetscCall(DMGetCoordinateDM(forest, &cdm));
972: PetscCall(DMGetLocalSection(cdm, &cs));
973: for (c = cStart; c < cEnd; c++) {
974: PetscInt doit = 0, outside = 0;
975: PetscCall(DMPlexVecGetClosure(cdm, cs, coords, c, &csize, &coef));
976: Nv = csize / dim;
977: for (nz = d = 0; d < Nv; d++) {
978: PetscReal z = PetscRealPart(coef[d * dim + (dim - 1)]), x = PetscSqr(PetscRealPart(coef[d * dim + 0])) + ((dim == 3) ? PetscSqr(PetscRealPart(coef[d * dim + 1])) : 0);
979: x = PetscSqrtReal(x);
980: if (type == 0) {
981: 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 */
982: } else if (type == 1 && (z > ctx->vperp0_radius1 || z < -ctx->vperp0_radius1)) {
983: outside++; /* don't refine outside electron refine radius */
984: PetscCall(PetscInfo(sol, "\t%" PetscInt_FMT ") (debug) found %s cells\n", grid, s_refine_names[type]));
985: } else if (type == 3 && (z > ctx->vperp0_radius2 || z < -ctx->vperp0_radius2)) {
986: outside++; /* refine r=0 cells on refinement front */
987: PetscCall(PetscInfo(sol, "\t%" PetscInt_FMT ") (debug) found %s cells\n", grid, s_refine_names[type]));
988: }
989: if (x < PETSC_MACHINE_EPSILON * 10. && (type != 0 || ctx->re_radius > PETSC_SQRT_MACHINE_EPSILON)) nz++;
990: }
991: PetscCall(DMPlexVecRestoreClosure(cdm, cs, coords, c, &csize, &coef));
992: if (doit || (outside < Nv && nz)) {
993: PetscCall(DMLabelSetValue(adaptLabel, c, DM_ADAPT_REFINE));
994: nrefined++;
995: }
996: }
997: PetscCall(PetscInfo(sol, "\t%" PetscInt_FMT ") Refined %" PetscInt_FMT " cells\n", grid, nrefined));
998: }
999: PetscCall(DMDestroy(&plex));
1000: PetscCall(DMAdaptLabel(forest, adaptLabel, &adaptedDM));
1001: PetscCall(DMLabelDestroy(&adaptLabel));
1002: *newForest = adaptedDM;
1003: if (adaptedDM) {
1004: if (isForest) PetscCall(DMForestSetAdaptivityForest(adaptedDM, NULL)); // ????
1005: PetscCall(DMConvert(adaptedDM, DMPLEX, &plex));
1006: PetscCall(DMPlexGetHeightStratum(plex, 0, &cStart, &cEnd));
1007: PetscCall(PetscInfo(sol, "\t\t\t\t%" PetscInt_FMT ") %" PetscInt_FMT " cells, %" PetscInt_FMT " total quadrature points\n", grid, cEnd - cStart, Nq * (cEnd - cStart)));
1008: PetscCall(DMDestroy(&plex));
1009: } else *newForest = NULL;
1010: PetscFunctionReturn(PETSC_SUCCESS);
1011: }
1013: // forest goes in (ctx->plex[grid]), plex comes out
1014: static PetscErrorCode adapt(PetscInt grid, LandauCtx *ctx, Vec *uu)
1015: {
1016: PetscInt adaptIter;
1018: PetscFunctionBegin;
1019: 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]};
1020: for (type = 0; type < 5; type++) {
1021: for (adaptIter = 0; adaptIter < limits[type]; adaptIter++) {
1022: DM newForest = NULL;
1023: PetscCall(adaptToleranceFEM(ctx->fe[0], *uu, type, grid, ctx, &newForest));
1024: if (newForest) {
1025: PetscCall(DMDestroy(&ctx->plex[grid]));
1026: PetscCall(VecDestroy(uu));
1027: PetscCall(DMCreateGlobalVector(newForest, uu));
1028: PetscCall(LandauSetInitialCondition(newForest, *uu, grid, 0, 1, ctx));
1029: ctx->plex[grid] = newForest;
1030: } else {
1031: PetscCall(PetscInfo(*uu, "No refinement\n"));
1032: }
1033: }
1034: }
1035: PetscCall(PetscObjectSetName((PetscObject)*uu, "uAMR"));
1036: PetscFunctionReturn(PETSC_SUCCESS);
1037: }
1039: // make log(Lambdas) from NRL Plasma formulary
1040: static PetscErrorCode makeLambdas(LandauCtx *ctx)
1041: {
1042: PetscFunctionBegin;
1043: for (PetscInt gridi = 0; gridi < ctx->num_grids; gridi++) {
1044: PetscInt iii = ctx->species_offset[gridi];
1045: PetscReal Ti_ev = (ctx->thermal_temps[iii] / 1.1604525e7) * 1000; // convert (back) to eV
1046: PetscReal ni = ctx->n[iii] * ctx->n_0;
1047: for (PetscInt gridj = gridi; gridj < ctx->num_grids; gridj++) {
1048: PetscInt jjj = ctx->species_offset[gridj];
1049: PetscReal Zj = ctx->charges[jjj] / 1.6022e-19;
1050: if (gridi == 0) {
1051: if (gridj == 0) { // lam_ee
1052: ctx->lambdas[gridi][gridj] = 23.5 - PetscLogReal(PetscSqrtReal(ni) * PetscPowReal(Ti_ev, -1.25)) - PetscSqrtReal(1e-5 + PetscSqr(PetscLogReal(Ti_ev) - 2) / 16);
1053: } else { // lam_ei == lam_ie
1054: if (10 * Zj * Zj > Ti_ev) {
1055: ctx->lambdas[gridi][gridj] = ctx->lambdas[gridj][gridi] = 23 - PetscLogReal(PetscSqrtReal(ni) * Zj * PetscPowReal(Ti_ev, -1.5));
1056: } else {
1057: ctx->lambdas[gridi][gridj] = ctx->lambdas[gridj][gridi] = 24 - PetscLogReal(PetscSqrtReal(ni) / Ti_ev);
1058: }
1059: }
1060: } else { // lam_ii'
1061: PetscReal mui = ctx->masses[iii] / 1.6720e-27, Zi = ctx->charges[iii] / 1.6022e-19;
1062: PetscReal Tj_ev = (ctx->thermal_temps[jjj] / 1.1604525e7) * 1000; // convert (back) to eV
1063: PetscReal muj = ctx->masses[jjj] / 1.6720e-27;
1064: PetscReal nj = ctx->n[jjj] * ctx->n_0;
1065: 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));
1066: }
1067: }
1068: }
1069: //PetscReal v0 = PetscSqrtReal(ctx->k * ctx->thermal_temps[iii] / ctx->masses[iii]); /* arbitrary units for non-dimensionalization: plasma formulary def */
1070: PetscFunctionReturn(PETSC_SUCCESS);
1071: }
1073: static PetscErrorCode ProcessOptions(LandauCtx *ctx, const char prefix[])
1074: {
1075: PetscBool flg, fileflg;
1076: PetscInt ii, nt, nm, nc, num_species_grid[LANDAU_MAX_GRIDS], non_dim_grid;
1077: PetscReal lnLam = 10;
1078: DM dummy;
1080: PetscFunctionBegin;
1081: PetscCall(DMCreate(ctx->comm, &dummy));
1082: /* get options - initialize context */
1083: ctx->verbose = 1; // should be 0 for silent compliance
1084: ctx->batch_sz = 1;
1085: ctx->batch_view_idx = 0;
1086: ctx->interpolate = PETSC_TRUE;
1087: ctx->gpu_assembly = PETSC_TRUE;
1088: ctx->norm_state = 0;
1089: ctx->electronShift = 0;
1090: ctx->M = NULL;
1091: ctx->J = NULL;
1092: /* geometry and grids */
1093: ctx->sphere = PETSC_FALSE;
1094: ctx->map_sphere = PETSC_TRUE;
1095: ctx->use_p4est = PETSC_FALSE;
1096: ctx->simplex = PETSC_FALSE;
1097: for (PetscInt grid = 0; grid < LANDAU_MAX_GRIDS; grid++) {
1098: ctx->radius[grid] = 5.; /* thermal radius (velocity) */
1099: ctx->radius_perp[grid] = 5.; /* thermal radius (velocity) */
1100: ctx->radius_par[grid] = 5.; /* thermal radius (velocity) */
1101: ctx->numAMRRefine[grid] = 0;
1102: ctx->postAMRRefine[grid] = 0;
1103: ctx->species_offset[grid + 1] = 1; // one species default
1104: num_species_grid[grid] = 0;
1105: ctx->plex[grid] = NULL; /* cache as expensive to Convert */
1106: }
1107: ctx->species_offset[0] = 0;
1108: ctx->re_radius = 0.;
1109: ctx->vperp0_radius1 = 0;
1110: ctx->vperp0_radius2 = 0;
1111: ctx->nZRefine1 = 0;
1112: ctx->nZRefine2 = 0;
1113: ctx->numRERefine = 0;
1114: num_species_grid[0] = 1; // one species default
1115: /* species - [0] electrons, [1] one ion species eg, duetarium, [2] heavy impurity ion, ... */
1116: ctx->charges[0] = -1; /* electron charge (MKS) */
1117: ctx->masses[0] = 1 / 1835.469965278441013; /* temporary value in proton mass */
1118: ctx->n[0] = 1;
1119: ctx->v_0 = 1; /* thermal velocity, we could start with a scale != 1 */
1120: ctx->thermal_temps[0] = 1;
1121: /* constants, etc. */
1122: ctx->epsilon0 = 8.8542e-12; /* permittivity of free space (MKS) F/m */
1123: ctx->k = 1.38064852e-23; /* Boltzmann constant (MKS) J/K */
1124: ctx->n_0 = 1.e20; /* typical plasma n, but could set it to 1 */
1125: ctx->Ez = 0;
1126: for (PetscInt grid = 0; grid < LANDAU_NUM_TIMERS; grid++) ctx->times[grid] = 0;
1127: for (PetscInt ii = 0; ii < LANDAU_DIM; ii++) ctx->cells0[ii] = 2;
1128: if (LANDAU_DIM == 2) ctx->cells0[0] = 1;
1129: ctx->use_matrix_mass = PETSC_FALSE;
1130: ctx->use_relativistic_corrections = PETSC_FALSE;
1131: ctx->use_energy_tensor_trick = PETSC_FALSE; /* Use Eero's trick for energy conservation v --> grad(v^2/2) */
1132: ctx->SData_d.w = NULL;
1133: ctx->SData_d.x = NULL;
1134: ctx->SData_d.y = NULL;
1135: ctx->SData_d.z = NULL;
1136: ctx->SData_d.invJ = NULL;
1137: ctx->jacobian_field_major_order = PETSC_FALSE;
1138: ctx->SData_d.coo_elem_offsets = NULL;
1139: ctx->SData_d.coo_elem_point_offsets = NULL;
1140: ctx->SData_d.coo_elem_fullNb = NULL;
1141: ctx->SData_d.coo_size = 0;
1142: PetscOptionsBegin(ctx->comm, prefix, "Options for Fokker-Plank-Landau collision operator", "none");
1143: #if defined(PETSC_HAVE_KOKKOS)
1144: ctx->deviceType = LANDAU_KOKKOS;
1145: PetscCall(PetscStrncpy(ctx->filename, "kokkos", sizeof(ctx->filename)));
1146: #else
1147: ctx->deviceType = LANDAU_CPU;
1148: PetscCall(PetscStrncpy(ctx->filename, "cpu", sizeof(ctx->filename)));
1149: #endif
1150: PetscCall(PetscOptionsString("-dm_landau_device_type", "Use kernels on 'cpu' 'kokkos'", "plexland.c", ctx->filename, ctx->filename, sizeof(ctx->filename), NULL));
1151: PetscCall(PetscStrcmp("cpu", ctx->filename, &flg));
1152: if (flg) {
1153: ctx->deviceType = LANDAU_CPU;
1154: } else {
1155: PetscCall(PetscStrcmp("kokkos", ctx->filename, &flg));
1156: PetscCheck(flg, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_device_type %s", ctx->filename);
1157: ctx->deviceType = LANDAU_KOKKOS;
1158: }
1159: ctx->filename[0] = '\0';
1160: PetscCall(PetscOptionsString("-dm_landau_filename", "file to read mesh from", "plexland.c", ctx->filename, ctx->filename, sizeof(ctx->filename), &fileflg));
1161: PetscCall(PetscOptionsReal("-dm_landau_electron_shift", "Shift in thermal velocity of electrons", "none", ctx->electronShift, &ctx->electronShift, NULL));
1162: PetscCall(PetscOptionsInt("-dm_landau_verbose", "Level of verbosity output", "plexland.c", ctx->verbose, &ctx->verbose, NULL));
1163: PetscCall(PetscOptionsInt("-dm_landau_batch_size", "Number of 'vertices' to batch", "ex2.c", ctx->batch_sz, &ctx->batch_sz, NULL));
1164: 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);
1165: 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));
1166: 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);
1167: PetscCall(PetscOptionsReal("-dm_landau_Ez", "Initial parallel electric field in unites of Conner-Hastie critical field", "plexland.c", ctx->Ez, &ctx->Ez, NULL));
1168: PetscCall(PetscOptionsReal("-dm_landau_n_0", "Normalization constant for number density", "plexland.c", ctx->n_0, &ctx->n_0, NULL));
1169: 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));
1170: PetscCall(PetscOptionsBool("-dm_landau_use_relativistic_corrections", "Use relativistic corrections", "plexland.c", ctx->use_relativistic_corrections, &ctx->use_relativistic_corrections, NULL));
1171: PetscCall(PetscOptionsBool("-dm_landau_simplex", "Use simplex elements", "plexland.c", ctx->simplex, &ctx->simplex, NULL));
1172: PetscCall(PetscOptionsBool("-dm_landau_sphere", "use sphere/semi-circle domain instead of rectangle", "plexland.c", ctx->sphere, &ctx->sphere, NULL));
1173: PetscCall(PetscOptionsBool("-dm_landau_map_sphere", "Map to sphere/semi-circle domain instead of rectangle", "plexland.c", ctx->map_sphere, &ctx->map_sphere, NULL));
1174: if (LANDAU_DIM == 2 && ctx->use_relativistic_corrections) ctx->use_relativistic_corrections = PETSC_FALSE; // should warn
1175: 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,
1176: &ctx->use_energy_tensor_trick, NULL));
1178: /* get num species with temperature, set defaults */
1179: for (ii = 1; ii < LANDAU_MAX_SPECIES; ii++) {
1180: ctx->thermal_temps[ii] = 1;
1181: ctx->charges[ii] = 1;
1182: ctx->masses[ii] = 1;
1183: ctx->n[ii] = 1;
1184: }
1185: nt = LANDAU_MAX_SPECIES;
1186: 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));
1187: PetscCheck(flg, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_thermal_temps ,t1,t2,.. must be provided to set the number of species");
1188: PetscCall(PetscInfo(dummy, "num_species set to number of thermal temps provided (%" PetscInt_FMT ")\n", nt));
1189: ctx->num_species = nt;
1190: for (ii = 0; ii < ctx->num_species; ii++) ctx->thermal_temps[ii] *= 1.1604525e7; /* convert to Kelvin */
1191: nm = LANDAU_MAX_SPECIES - 1;
1192: 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));
1193: 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);
1194: nm = LANDAU_MAX_SPECIES;
1195: PetscCall(PetscOptionsRealArray("-dm_landau_n", "Number density of each species = n_s * n_0", "plexland.c", ctx->n, &nm, &flg));
1196: 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);
1197: for (ii = 0; ii < LANDAU_MAX_SPECIES; ii++) ctx->masses[ii] *= 1.6720e-27; /* scale by proton mass kg */
1198: ctx->masses[0] = 9.10938356e-31; /* electron mass kg (should be about right already) */
1199: nc = LANDAU_MAX_SPECIES - 1;
1200: 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));
1201: 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);
1202: for (ii = 0; ii < LANDAU_MAX_SPECIES; ii++) ctx->charges[ii] *= 1.6022e-19; /* electron/proton charge (MKS) */
1203: /* geometry and grids */
1204: nt = LANDAU_MAX_GRIDS;
1205: 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));
1206: if (flg) {
1207: ctx->num_grids = nt;
1208: for (ii = nt = 0; ii < ctx->num_grids; ii++) nt += num_species_grid[ii];
1209: 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,
1210: ctx->num_grids, LANDAU_MAX_GRIDS);
1211: } else {
1212: if (ctx->num_species > LANDAU_MAX_GRIDS) {
1213: num_species_grid[0] = 1;
1214: num_species_grid[1] = ctx->num_species - 1;
1215: ctx->num_grids = 2;
1216: } else {
1217: ctx->num_grids = ctx->num_species;
1218: for (ii = 0; ii < ctx->num_grids; ii++) num_species_grid[ii] = 1;
1219: }
1220: }
1221: 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];
1222: 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],
1223: ctx->num_species);
1224: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1225: PetscInt iii = ctx->species_offset[grid]; // normalize with first (arbitrary) species on grid
1226: ctx->thermal_speed[grid] = PetscSqrtReal(ctx->k * ctx->thermal_temps[iii] / ctx->masses[iii]); /* arbitrary units for non-dimensionalization: plasma formulary def */
1227: }
1228: // get lambdas here because we need them for t_0 etc
1229: PetscCall(PetscOptionsReal("-dm_landau_ln_lambda", "Universal cross section parameter. Default uses NRL formulas", "plexland.c", lnLam, &lnLam, &flg));
1230: if (flg) {
1231: for (PetscInt grid = 0; grid < LANDAU_MAX_GRIDS; grid++) {
1232: for (PetscInt gridj = 0; gridj < LANDAU_MAX_GRIDS; gridj++) ctx->lambdas[gridj][grid] = lnLam; /* cross section ratio large - small angle collisions */
1233: }
1234: } else {
1235: PetscCall(makeLambdas(ctx));
1236: }
1237: non_dim_grid = 0;
1238: 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));
1239: if (non_dim_grid != 0) PetscCall(PetscInfo(dummy, "Normalization grid set to %" PetscInt_FMT ", but non-default not well verified\n", non_dim_grid));
1240: 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);
1241: ctx->v_0 = ctx->thermal_speed[non_dim_grid]; /* arbitrary units for non dimensionalization: global mean velocity in 1D of electrons */
1242: ctx->m_0 = ctx->masses[non_dim_grid]; /* arbitrary reference mass, electrons */
1243: 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 */
1244: /* domain */
1245: nt = LANDAU_MAX_GRIDS;
1246: PetscCall(PetscOptionsRealArray("-dm_landau_domain_radius", "Phase space size in units of thermal velocity of grid", "plexland.c", ctx->radius, &nt, &flg));
1247: if (flg) {
1248: 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);
1249: while (nt--) ctx->radius_par[nt] = ctx->radius_perp[nt] = ctx->radius[nt];
1250: } else {
1251: nt = LANDAU_MAX_GRIDS;
1252: 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));
1253: 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);
1254: 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));
1255: 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);
1256: }
1257: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1258: if (flg && ctx->radius[grid] <= 0) { /* negative is ratio of c - need to set par and perp with this -- todo */
1259: if (ctx->radius[grid] == 0) ctx->radius[grid] = 0.75;
1260: else ctx->radius[grid] = -ctx->radius[grid];
1261: 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)
1262: PetscCall(PetscInfo(dummy, "Change domain radius to %g for grid %" PetscInt_FMT "\n", (double)ctx->radius[grid], grid));
1263: }
1264: ctx->radius[grid] *= ctx->thermal_speed[grid] / ctx->v_0; // scale domain by thermal radius relative to v_0
1265: ctx->radius_perp[grid] *= ctx->thermal_speed[grid] / ctx->v_0; // scale domain by thermal radius relative to v_0
1266: ctx->radius_par[grid] *= ctx->thermal_speed[grid] / ctx->v_0; // scale domain by thermal radius relative to v_0
1267: }
1268: /* amr parameters */
1269: if (!fileflg) {
1270: nt = LANDAU_MAX_GRIDS;
1271: 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));
1272: 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);
1273: nt = LANDAU_MAX_GRIDS;
1274: PetscCall(PetscOptionsIntArray("-dm_landau_amr_post_refine", "Number of levels to uniformly refine after AMR", "plexland.c", ctx->postAMRRefine, &nt, &flg));
1275: for (ii = 1; ii < ctx->num_grids; ii++) ctx->postAMRRefine[ii] = ctx->postAMRRefine[0]; // all grids the same now
1276: 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));
1277: 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));
1278: 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));
1279: 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));
1280: 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));
1281: 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));
1282: /* spherical domain */
1283: if (ctx->sphere || ctx->simplex) {
1284: ctx->sphere_uniform_normal = PETSC_FALSE;
1285: 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));
1286: if (!ctx->sphere_uniform_normal) { // true
1287: nt = LANDAU_MAX_GRIDS;
1288: 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));
1289: if (flg && nt < ctx->num_grids) {
1290: for (PetscInt grid = nt; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_90degree[grid] = ctx->sphere_inner_radius_90degree[0];
1291: } else if (!flg || nt == 0) {
1292: if (ctx->sphere && !ctx->simplex && LANDAU_DIM == 3) {
1293: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_90degree[grid] = 0.35; // optimized for R=6, Q4, AMR=0, 0 refinement
1294: } else {
1295: if (LANDAU_DIM == 2) {
1296: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_90degree[grid] = 0.4; // optimized for R=5, Q4, AMR=0
1297: } else {
1298: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_90degree[grid] = 0.577 * 0.40;
1299: }
1300: }
1301: }
1302: nt = LANDAU_MAX_GRIDS;
1303: 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));
1304: if (flg && nt < ctx->num_grids) {
1305: for (PetscInt grid = nt; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_45degree[grid] = ctx->sphere_inner_radius_45degree[0];
1306: } else if (!flg || nt == 0) {
1307: if (LANDAU_DIM == 2) {
1308: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_45degree[grid] = 0.45; // optimized for R=5, Q4, AMR=0
1309: } else {
1310: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_45degree[grid] = 0.4; // 3D sphere
1311: }
1312: }
1313: 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]));
1314: } else {
1315: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1316: switch (ctx->numAMRRefine[grid]) {
1317: case 0:
1318: case 1:
1319: case 2:
1320: case 3:
1321: default:
1322: if (LANDAU_DIM == 2) {
1323: ctx->sphere_inner_radius_90degree[grid] = 0.40;
1324: ctx->sphere_inner_radius_45degree[grid] = 0.45;
1325: } else {
1326: ctx->sphere_inner_radius_45degree[grid] = 0.25;
1327: }
1328: }
1329: }
1330: }
1331: } else {
1332: nt = LANDAU_DIM;
1333: PetscCall(PetscOptionsIntArray("-dm_landau_num_cells", "Number of cells in each dimension of base grid", "plexland.c", ctx->cells0, &nt, &flg));
1334: }
1335: }
1336: /* processing options */
1337: PetscCall(PetscOptionsBool("-dm_landau_gpu_assembly", "Assemble Jacobian on GPU", "plexland.c", ctx->gpu_assembly, &ctx->gpu_assembly, NULL));
1338: 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));
1339: 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");
1340: PetscCheck(!ctx->jacobian_field_major_order, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_jacobian_field_major_order DEPRECATED");
1341: PetscOptionsEnd();
1343: for (ii = ctx->num_species; ii < LANDAU_MAX_SPECIES; ii++) ctx->masses[ii] = ctx->thermal_temps[ii] = ctx->charges[ii] = 0;
1344: if (ctx->verbose != 0) {
1345: PetscReal pmassunit = PetscRealConstant(1.6720e-27);
1347: 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)));
1348: 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)));
1349: 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)));
1350: 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],
1351: (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));
1352: 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]));
1353: 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]));
1354: if (ctx->use_relativistic_corrections) PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\nUse relativistic corrections\n"));
1355: else PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
1356: }
1357: PetscCall(DMDestroy(&dummy));
1358: {
1359: PetscMPIInt rank;
1360: PetscCallMPI(MPI_Comm_rank(PETSC_COMM_WORLD, &rank));
1361: ctx->stage = 0;
1362: PetscCall(PetscLogEventRegister("Landau Create", DM_CLASSID, &ctx->events[13])); /* 13 */
1363: PetscCall(PetscLogEventRegister(" GPU ass. setup", DM_CLASSID, &ctx->events[2])); /* 2 */
1364: PetscCall(PetscLogEventRegister(" Build matrix", DM_CLASSID, &ctx->events[12])); /* 12 */
1365: PetscCall(PetscLogEventRegister(" Assembly maps", DM_CLASSID, &ctx->events[15])); /* 15 */
1366: PetscCall(PetscLogEventRegister("Landau Mass mat", DM_CLASSID, &ctx->events[14])); /* 14 */
1367: PetscCall(PetscLogEventRegister("Landau Operator", DM_CLASSID, &ctx->events[11])); /* 11 */
1368: PetscCall(PetscLogEventRegister("Landau Jacobian", DM_CLASSID, &ctx->events[0])); /* 0 */
1369: PetscCall(PetscLogEventRegister("Landau Mass", DM_CLASSID, &ctx->events[9])); /* 9 */
1370: PetscCall(PetscLogEventRegister(" Preamble", DM_CLASSID, &ctx->events[10])); /* 10 */
1371: PetscCall(PetscLogEventRegister(" static IP Data", DM_CLASSID, &ctx->events[7])); /* 7 */
1372: PetscCall(PetscLogEventRegister(" dynamic IP-Jac", DM_CLASSID, &ctx->events[1])); /* 1 */
1373: PetscCall(PetscLogEventRegister(" Kernel-init", DM_CLASSID, &ctx->events[3])); /* 3 */
1374: PetscCall(PetscLogEventRegister(" Jac-f-df (GPU)", DM_CLASSID, &ctx->events[8])); /* 8 */
1375: PetscCall(PetscLogEventRegister(" J Kernel (GPU)", DM_CLASSID, &ctx->events[4])); /* 4 */
1376: PetscCall(PetscLogEventRegister(" M Kernel (GPU)", DM_CLASSID, &ctx->events[16])); /* 16 */
1377: PetscCall(PetscLogEventRegister(" Copy to CPU", DM_CLASSID, &ctx->events[5])); /* 5 */
1378: PetscCall(PetscLogEventRegister(" CPU assemble", DM_CLASSID, &ctx->events[6])); /* 6 */
1380: if (rank) { /* turn off output stuff for duplicate runs - do we need to add the prefix to all this? */
1381: PetscCall(PetscOptionsClearValue(NULL, "-snes_converged_reason"));
1382: PetscCall(PetscOptionsClearValue(NULL, "-ksp_converged_reason"));
1383: PetscCall(PetscOptionsClearValue(NULL, "-snes_monitor"));
1384: PetscCall(PetscOptionsClearValue(NULL, "-ksp_monitor"));
1385: PetscCall(PetscOptionsClearValue(NULL, "-ts_monitor"));
1386: PetscCall(PetscOptionsClearValue(NULL, "-ts_view"));
1387: PetscCall(PetscOptionsClearValue(NULL, "-ts_adapt_monitor"));
1388: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_amr_dm_view"));
1389: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_amr_vec_view"));
1390: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_mass_dm_view"));
1391: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_mass_view"));
1392: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_jacobian_view"));
1393: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_mat_view"));
1394: PetscCall(PetscOptionsClearValue(NULL, "-pc_bjkokkos_ksp_converged_reason"));
1395: PetscCall(PetscOptionsClearValue(NULL, "-pc_bjkokkos_ksp_monitor"));
1396: PetscCall(PetscOptionsClearValue(NULL, "-"));
1397: PetscCall(PetscOptionsClearValue(NULL, "-info"));
1398: }
1399: }
1400: PetscFunctionReturn(PETSC_SUCCESS);
1401: }
1403: static PetscErrorCode CreateStaticData(PetscInt dim, IS grid_batch_is_inv[], const char prefix[], LandauCtx *ctx)
1404: {
1405: PetscSection section[LANDAU_MAX_GRIDS], globsection[LANDAU_MAX_GRIDS];
1406: PetscQuadrature quad;
1407: const PetscReal *quadWeights;
1408: PetscReal invMass[LANDAU_MAX_SPECIES], nu_alpha[LANDAU_MAX_SPECIES], nu_beta[LANDAU_MAX_SPECIES];
1409: 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;
1410: PetscTabulation *Tf;
1411: PetscDS prob;
1413: PetscFunctionBegin;
1414: PetscCall(PetscFEGetDimension(ctx->fe[0], &Nb));
1415: 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);
1416: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1417: for (PetscInt ii = ctx->species_offset[grid]; ii < ctx->species_offset[grid + 1]; ii++) {
1418: invMass[ii] = ctx->m_0 / ctx->masses[ii];
1419: nu_alpha[ii] = PetscSqr(ctx->charges[ii] / ctx->m_0) * ctx->m_0 / ctx->masses[ii];
1420: nu_beta[ii] = PetscSqr(ctx->charges[ii] / ctx->epsilon0) / (8 * PETSC_PI) * ctx->t_0 * ctx->n_0 / PetscPowReal(ctx->v_0, 3);
1421: }
1422: }
1423: if (ctx->verbose == 4) {
1424: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "nu_alpha: "));
1425: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1426: PetscInt iii = ctx->species_offset[grid];
1427: for (PetscInt ii = iii; ii < ctx->species_offset[grid + 1]; ii++) PetscCall(PetscPrintf(PETSC_COMM_WORLD, " %e", (double)nu_alpha[ii]));
1428: }
1429: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\nnu_beta: "));
1430: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1431: PetscInt iii = ctx->species_offset[grid];
1432: for (PetscInt ii = iii; ii < ctx->species_offset[grid + 1]; ii++) PetscCall(PetscPrintf(PETSC_COMM_WORLD, " %e", (double)nu_beta[ii]));
1433: }
1434: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\nnu_alpha[i]*nu_beta[j]*lambda[i][j]:\n"));
1435: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1436: PetscInt iii = ctx->species_offset[grid];
1437: for (PetscInt ii = iii; ii < ctx->species_offset[grid + 1]; ii++) {
1438: for (PetscInt gridj = 0; gridj < ctx->num_grids; gridj++) {
1439: PetscInt jjj = ctx->species_offset[gridj];
1440: 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])));
1441: }
1442: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
1443: }
1444: }
1445: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "lambda[i][j]:\n"));
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++) {
1449: for (PetscInt gridj = 0; gridj < ctx->num_grids; gridj++) {
1450: PetscInt jjj = ctx->species_offset[gridj];
1451: for (PetscInt jj = jjj; jj < ctx->species_offset[gridj + 1]; jj++) PetscCall(PetscPrintf(PETSC_COMM_WORLD, " %14.9e", (double)ctx->lambdas[grid][gridj]));
1452: }
1453: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
1454: }
1455: }
1456: }
1457: PetscCall(DMGetDS(ctx->plex[0], &prob)); // same DS for all grids
1458: PetscCall(PetscDSGetTabulation(prob, &Tf)); // Bf, &Df same for all grids
1459: /* DS, Tab and quad is same on all grids */
1460: PetscCheck(ctx->plex[0], ctx->comm, PETSC_ERR_ARG_WRONG, "Plex not created");
1461: PetscCall(PetscFEGetQuadrature(ctx->fe[0], &quad));
1462: PetscCall(PetscQuadratureGetData(quad, NULL, NULL, &Nq, NULL, &quadWeights));
1463: 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);
1464: /* setup each grid */
1465: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1466: PetscInt cStart, cEnd;
1467: PetscCheck(ctx->plex[grid] != NULL, ctx->comm, PETSC_ERR_ARG_WRONG, "Plex not created");
1468: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd));
1469: numCells[grid] = cEnd - cStart; // grids can have different topology
1470: PetscCall(DMGetLocalSection(ctx->plex[grid], §ion[grid]));
1471: PetscCall(DMGetGlobalSection(ctx->plex[grid], &globsection[grid]));
1472: PetscCall(PetscSectionGetNumFields(section[grid], &Nf[grid]));
1473: ncellsTot += numCells[grid];
1474: }
1475: /* create GPU assembly data */
1476: if (ctx->gpu_assembly) { /* we need GPU object with GPU assembly */
1477: PetscContainer container;
1478: PetscScalar *elemMatrix, *elMat;
1479: pointInterpolationP4est(*pointMaps)[LANDAU_MAX_Q_FACE];
1480: P4estVertexMaps *maps;
1481: const PetscInt *plex_batch = NULL, elMatSz = Nb * Nb * ctx->num_species * ctx->num_species;
1482: LandauIdx *coo_elem_offsets = NULL, *coo_elem_fullNb = NULL, (*coo_elem_point_offsets)[LANDAU_MAX_NQND + 1] = NULL;
1483: /* create GPU assembly data */
1484: PetscCall(PetscInfo(ctx->plex[0], "Make GPU maps %d\n", 1));
1485: PetscCall(PetscLogEventBegin(ctx->events[2], 0, 0, 0, 0));
1486: PetscCall(PetscMalloc(sizeof(*maps) * ctx->num_grids, &maps));
1487: PetscCall(PetscMalloc(sizeof(*pointMaps) * MAP_BF_SIZE, &pointMaps));
1488: PetscCall(PetscMalloc(sizeof(*elemMatrix) * elMatSz, &elemMatrix));
1490: { // setup COO assembly -- put COO metadata directly in ctx->SData_d
1491: PetscCall(PetscMalloc3(ncellsTot + 1, &coo_elem_offsets, ncellsTot, &coo_elem_fullNb, ncellsTot, &coo_elem_point_offsets)); // array of integer pointers
1492: coo_elem_offsets[0] = 0; // finish later
1493: PetscCall(PetscInfo(ctx->plex[0], "COO initialization, %" PetscInt_FMT " cells\n", ncellsTot));
1494: ctx->SData_d.coo_n_cellsTot = ncellsTot;
1495: ctx->SData_d.coo_elem_offsets = (void *)coo_elem_offsets;
1496: ctx->SData_d.coo_elem_fullNb = (void *)coo_elem_fullNb;
1497: ctx->SData_d.coo_elem_point_offsets = (void *)coo_elem_point_offsets;
1498: }
1500: ctx->SData_d.coo_max_fullnb = 0;
1501: for (PetscInt grid = 0, glb_elem_idx = 0; grid < ctx->num_grids; grid++) {
1502: PetscInt cStart, cEnd, Nfloc = Nf[grid], totDim = Nfloc * Nb;
1503: if (grid_batch_is_inv[grid]) PetscCall(ISGetIndices(grid_batch_is_inv[grid], &plex_batch));
1504: PetscCheck(!plex_batch, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_jacobian_field_major_order DEPRECATED");
1505: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd));
1506: // make maps
1507: maps[grid].d_self = NULL;
1508: maps[grid].num_elements = numCells[grid];
1509: maps[grid].num_face = (PetscInt)(pow(Nq, 1. / ((double)dim)) + .001); // Q
1510: maps[grid].num_face = (PetscInt)(pow(maps[grid].num_face, (double)(dim - 1)) + .001); // Q^2
1511: maps[grid].num_reduced = 0;
1512: maps[grid].deviceType = ctx->deviceType;
1513: maps[grid].numgrids = ctx->num_grids;
1514: // count reduced and get
1515: PetscCall(PetscMalloc(maps[grid].num_elements * sizeof(*maps[grid].gIdx), &maps[grid].gIdx));
1516: for (PetscInt ej = cStart, eidx = 0; ej < cEnd; ++ej, ++eidx, glb_elem_idx++) {
1517: if (coo_elem_offsets) coo_elem_offsets[glb_elem_idx + 1] = coo_elem_offsets[glb_elem_idx]; // start with last one, then add
1518: for (PetscInt fieldA = 0; fieldA < Nf[grid]; fieldA++) {
1519: PetscInt fullNb = 0;
1520: for (PetscInt q = 0; q < Nb; ++q) {
1521: PetscInt numindices, *indices;
1522: PetscScalar *valuesOrig = elMat = elemMatrix;
1523: PetscCall(PetscArrayzero(elMat, totDim * totDim));
1524: elMat[(fieldA * Nb + q) * totDim + fieldA * Nb + q] = 1;
1525: PetscCall(DMPlexGetClosureIndices(ctx->plex[grid], section[grid], globsection[grid], ej, PETSC_TRUE, &numindices, &indices, NULL, &elMat));
1526: if (ctx->simplex) {
1527: PetscCheck(numindices == Nb, ctx->comm, PETSC_ERR_ARG_WRONG, "numindices != Nb numindices=%" PetscInt_FMT " Nb=%" PetscInt_FMT, numindices, Nb);
1528: for (PetscInt q = 0; q < numindices; ++q) maps[grid].gIdx[eidx][fieldA][q] = indices[q];
1529: fullNb++;
1530: } else {
1531: for (PetscInt f = 0; f < numindices; ++f) { // look for a non-zero on the diagonal (is this too complicated for simplices?)
1532: if (PetscAbs(PetscRealPart(elMat[f * numindices + f])) > PETSC_MACHINE_EPSILON) {
1533: // found it
1534: if (PetscAbs(PetscRealPart(elMat[f * numindices + f] - 1.)) < PETSC_MACHINE_EPSILON) { // normal vertex 1.0
1535: if (plex_batch) {
1536: maps[grid].gIdx[eidx][fieldA][q] = plex_batch[indices[f]];
1537: } else {
1538: maps[grid].gIdx[eidx][fieldA][q] = indices[f];
1539: }
1540: fullNb++;
1541: } else { //found a constraint
1542: PetscInt jj = 0;
1543: PetscReal sum = 0;
1544: const PetscInt ff = f;
1545: maps[grid].gIdx[eidx][fieldA][q] = -maps[grid].num_reduced - 1; // store (-)index: id = -(idx+1): idx = -id - 1
1546: PetscCheck(!ctx->simplex, ctx->comm, PETSC_ERR_ARG_WRONG, "No constraints with simplex");
1547: do { // constraints are continuous in Plex - exploit that here
1548: PetscInt ii; // get 'scale'
1549: 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
1550: 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
1551: pointMaps[maps[grid].num_reduced][jj].scale += PetscRealPart(elMat[f * numindices + ff + ii]);
1552: }
1553: }
1554: sum += pointMaps[maps[grid].num_reduced][jj].scale; // diagnostic
1555: // get 'gid'
1556: if (pointMaps[maps[grid].num_reduced][jj].scale == 0) pointMaps[maps[grid].num_reduced][jj].gid = -1; // 3D has Q and Q^2 interps
1557: else {
1558: if (plex_batch) {
1559: pointMaps[maps[grid].num_reduced][jj].gid = plex_batch[indices[f]];
1560: } else {
1561: pointMaps[maps[grid].num_reduced][jj].gid = indices[f];
1562: }
1563: fullNb++;
1564: }
1565: } while (++jj < maps[grid].num_face && ++f < numindices); // jj is incremented if we hit the end
1566: while (jj < maps[grid].num_face) {
1567: pointMaps[maps[grid].num_reduced][jj].scale = 0;
1568: pointMaps[maps[grid].num_reduced][jj].gid = -1;
1569: jj++;
1570: }
1571: if (PetscAbs(sum - 1.0) > 10 * PETSC_MACHINE_EPSILON) { // debug
1572: PetscInt d, f;
1573: PetscReal tmp = 0;
1574: PetscCall(
1575: 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));
1576: for (d = 0, tmp = 0; d < numindices; ++d) {
1577: 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]));
1578: for (f = 0; f < numindices; ++f) tmp += PetscRealPart(elMat[d * numindices + f]);
1579: if (tmp != 0) PetscCall(PetscPrintf(ctx->comm, " | %22.16e\n", (double)tmp));
1580: }
1581: }
1582: maps[grid].num_reduced++;
1583: 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);
1584: }
1585: break;
1586: }
1587: }
1588: } // !simplex
1589: // cleanup
1590: PetscCall(DMPlexRestoreClosureIndices(ctx->plex[grid], section[grid], globsection[grid], ej, PETSC_TRUE, &numindices, &indices, NULL, &elMat));
1591: if (elMat != valuesOrig) PetscCall(DMRestoreWorkArray(ctx->plex[grid], numindices * numindices, MPIU_SCALAR, &elMat));
1592: }
1593: { // setup COO assembly
1594: coo_elem_offsets[glb_elem_idx + 1] += fullNb * fullNb; // one species block, adds a block for each species, on this element in this grid
1595: if (fieldA == 0) { // cache full Nb for this element, on this grid per species
1596: coo_elem_fullNb[glb_elem_idx] = fullNb;
1597: if (fullNb > ctx->SData_d.coo_max_fullnb) ctx->SData_d.coo_max_fullnb = fullNb;
1598: } 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);
1599: }
1600: } // field
1601: } // cell
1602: // allocate and copy point data maps[grid].gIdx[eidx][field][q]
1603: PetscCall(PetscMalloc(maps[grid].num_reduced * sizeof(*maps[grid].c_maps), &maps[grid].c_maps));
1604: for (PetscInt ej = 0; ej < maps[grid].num_reduced; ++ej) {
1605: for (PetscInt q = 0; q < maps[grid].num_face; ++q) {
1606: maps[grid].c_maps[ej][q].scale = pointMaps[ej][q].scale;
1607: maps[grid].c_maps[ej][q].gid = pointMaps[ej][q].gid;
1608: }
1609: }
1610: #if defined(PETSC_HAVE_KOKKOS)
1611: if (ctx->deviceType == LANDAU_KOKKOS) PetscCall(LandauKokkosCreateMatMaps(maps, pointMaps, Nf, grid)); // implies Kokkos does
1612: #endif
1613: if (plex_batch) {
1614: PetscCall(ISRestoreIndices(grid_batch_is_inv[grid], &plex_batch));
1615: PetscCall(ISDestroy(&grid_batch_is_inv[grid])); // we are done with this
1616: }
1617: } /* grids */
1618: // finish COO
1619: { // setup COO assembly
1620: PetscInt *oor, *ooc;
1621: ctx->SData_d.coo_size = coo_elem_offsets[ncellsTot] * ctx->batch_sz;
1622: PetscCall(PetscMalloc2(ctx->SData_d.coo_size, &oor, ctx->SData_d.coo_size, &ooc));
1623: for (PetscInt i = 0; i < ctx->SData_d.coo_size; i++) oor[i] = ooc[i] = -1;
1624: // get
1625: for (PetscInt grid = 0, glb_elem_idx = 0; grid < ctx->num_grids; grid++) {
1626: for (PetscInt ej = 0; ej < numCells[grid]; ++ej, glb_elem_idx++) {
1627: const PetscInt fullNb = coo_elem_fullNb[glb_elem_idx];
1628: 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
1629: coo_elem_point_offsets[glb_elem_idx][0] = 0;
1630: for (PetscInt f = 0, cnt2 = 0; f < Nb; f++) {
1631: PetscInt idx = Idxs[f];
1632: coo_elem_point_offsets[glb_elem_idx][f + 1] = coo_elem_point_offsets[glb_elem_idx][f]; // start at last
1633: if (idx >= 0) {
1634: cnt2++;
1635: coo_elem_point_offsets[glb_elem_idx][f + 1]++; // inc
1636: } else {
1637: idx = -idx - 1;
1638: for (PetscInt q = 0; q < maps[grid].num_face; q++) {
1639: if (maps[grid].c_maps[idx][q].gid < 0) break;
1640: cnt2++;
1641: coo_elem_point_offsets[glb_elem_idx][f + 1]++; // inc
1642: }
1643: }
1644: PetscCheck(cnt2 <= fullNb, PETSC_COMM_SELF, PETSC_ERR_PLIB, "wrong count %" PetscInt_FMT " < %" PetscInt_FMT, fullNb, cnt2);
1645: }
1646: 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);
1647: }
1648: }
1649: // set
1650: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
1651: for (PetscInt grid = 0, glb_elem_idx = 0; grid < ctx->num_grids; grid++) {
1652: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
1653: for (PetscInt ej = 0; ej < numCells[grid]; ++ej, glb_elem_idx++) {
1654: const PetscInt fullNb = coo_elem_fullNb[glb_elem_idx], fullNb2 = fullNb * fullNb;
1655: // set (i,j)
1656: for (PetscInt fieldA = 0; fieldA < Nf[grid]; fieldA++) {
1657: const LandauIdx *const Idxs = &maps[grid].gIdx[ej][fieldA][0];
1658: PetscInt rows[LANDAU_MAX_Q_FACE], cols[LANDAU_MAX_Q_FACE];
1659: for (PetscInt f = 0; f < Nb; ++f) {
1660: const PetscInt nr = coo_elem_point_offsets[glb_elem_idx][f + 1] - coo_elem_point_offsets[glb_elem_idx][f];
1661: if (nr == 1) rows[0] = Idxs[f];
1662: else {
1663: const PetscInt idx = -Idxs[f] - 1;
1664: for (PetscInt q = 0; q < nr; q++) rows[q] = maps[grid].c_maps[idx][q].gid;
1665: }
1666: for (PetscInt g = 0; g < Nb; ++g) {
1667: const PetscInt nc = coo_elem_point_offsets[glb_elem_idx][g + 1] - coo_elem_point_offsets[glb_elem_idx][g];
1668: if (nc == 1) cols[0] = Idxs[g];
1669: else {
1670: const PetscInt idx = -Idxs[g] - 1;
1671: for (PetscInt q = 0; q < nc; q++) cols[q] = maps[grid].c_maps[idx][q].gid;
1672: }
1673: 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];
1674: for (PetscInt q = 0, idx = idx0; q < nr; q++) {
1675: for (PetscInt d = 0; d < nc; d++, idx++) {
1676: oor[idx] = rows[q] + moffset;
1677: ooc[idx] = cols[d] + moffset;
1678: }
1679: }
1680: }
1681: }
1682: }
1683: } // cell
1684: } // grid
1685: } // batch
1686: PetscCall(MatSetPreallocationCOO(ctx->J, ctx->SData_d.coo_size, oor, ooc));
1687: PetscCall(PetscFree2(oor, ooc));
1688: }
1689: PetscCall(PetscFree(pointMaps));
1690: PetscCall(PetscFree(elemMatrix));
1691: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
1692: PetscCall(PetscContainerSetPointer(container, (void *)maps));
1693: PetscCall(PetscContainerSetCtxDestroy(container, LandauGPUMapsDestroy));
1694: PetscCall(PetscObjectCompose((PetscObject)ctx->J, "assembly_maps", (PetscObject)container));
1695: PetscCall(PetscContainerDestroy(&container));
1696: PetscCall(PetscLogEventEnd(ctx->events[2], 0, 0, 0, 0));
1697: } // end GPU assembly
1698: { /* create static point data, Jacobian called first, only one vertex copy */
1699: PetscReal *invJe, *ww, *xx, *yy, *zz = NULL, *invJ_a;
1700: PetscInt outer_ipidx, outer_ej, grid, nip_glb = 0;
1701: PetscFE fe;
1702: PetscCall(PetscLogEventBegin(ctx->events[7], 0, 0, 0, 0));
1703: PetscCall(PetscInfo(ctx->plex[0], "Initialize static data\n"));
1704: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) nip_glb += Nq * numCells[grid];
1705: /* collect f data, first time is for Jacobian, but make mass now */
1706: if (ctx->verbose != 0) {
1707: PetscInt ncells = 0, N;
1708: MatInfo info;
1709: PetscCall(MatGetInfo(ctx->J, MAT_LOCAL, &info));
1710: PetscCall(MatGetSize(ctx->J, &N, NULL));
1711: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ncells += numCells[grid];
1712: 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,
1713: ctx->num_species, Nb, dim, N, (PetscInt)info.nz_used));
1714: }
1715: PetscCall(PetscMalloc4(nip_glb, &ww, nip_glb, &xx, nip_glb, &yy, nip_glb * dim * dim, &invJ_a));
1716: if (dim == 3) PetscCall(PetscMalloc1(nip_glb, &zz));
1717: if (ctx->use_energy_tensor_trick) {
1718: PetscCall(PetscFECreateDefault(PETSC_COMM_SELF, dim, 1, ctx->simplex, prefix, PETSC_DECIDE, &fe));
1719: PetscCall(PetscObjectSetName((PetscObject)fe, "energy"));
1720: }
1721: /* init each grids static data - no batch */
1722: for (grid = 0, outer_ipidx = 0, outer_ej = 0; grid < ctx->num_grids; grid++) { // OpenMP (once)
1723: Vec v2_2 = NULL; // projected function: v^2/2 for non-relativistic, gamma... for relativistic
1724: PetscSection e_section;
1725: DM dmEnergy;
1726: PetscInt cStart, cEnd, ej;
1728: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd));
1729: // prep energy trick, get v^2 / 2 vector
1730: if (ctx->use_energy_tensor_trick) {
1731: PetscErrorCode (*energyf[1])(PetscInt, PetscReal, const PetscReal[], PetscInt, PetscScalar[], void *) = {ctx->use_relativistic_corrections ? gamma_m1_f : energy_f};
1732: Vec glob_v2;
1733: PetscReal *c2_0[1], data[1] = {PetscSqr(C_0(ctx->v_0))};
1735: PetscCall(DMClone(ctx->plex[grid], &dmEnergy));
1736: PetscCall(PetscObjectSetName((PetscObject)dmEnergy, "energy"));
1737: PetscCall(DMSetField(dmEnergy, 0, NULL, (PetscObject)fe));
1738: PetscCall(DMCreateDS(dmEnergy));
1739: PetscCall(DMGetLocalSection(dmEnergy, &e_section));
1740: PetscCall(DMGetGlobalVector(dmEnergy, &glob_v2));
1741: PetscCall(PetscObjectSetName((PetscObject)glob_v2, "trick"));
1742: c2_0[0] = &data[0];
1743: PetscCall(DMProjectFunction(dmEnergy, 0., energyf, (void **)c2_0, INSERT_ALL_VALUES, glob_v2));
1744: PetscCall(DMGetLocalVector(dmEnergy, &v2_2));
1745: PetscCall(VecZeroEntries(v2_2)); /* zero BCs so don't set */
1746: PetscCall(DMGlobalToLocalBegin(dmEnergy, glob_v2, INSERT_VALUES, v2_2));
1747: PetscCall(DMGlobalToLocalEnd(dmEnergy, glob_v2, INSERT_VALUES, v2_2));
1748: PetscCall(DMViewFromOptions(dmEnergy, NULL, "-energy_dm_view"));
1749: PetscCall(VecViewFromOptions(glob_v2, NULL, "-energy_vec_view"));
1750: PetscCall(DMRestoreGlobalVector(dmEnergy, &glob_v2));
1751: }
1752: /* append part of the IP data for each grid */
1753: for (ej = 0; ej < numCells[grid]; ++ej, ++outer_ej) {
1754: PetscScalar *coefs = NULL;
1755: 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);
1756: invJe = invJ_a + outer_ej * Nq * dim * dim;
1757: PetscCall(DMPlexComputeCellGeometryFEM(ctx->plex[grid], ej + cStart, quad, vj, Jdummy, invJe, detJj));
1758: if (ctx->use_energy_tensor_trick) PetscCall(DMPlexVecGetClosure(dmEnergy, e_section, v2_2, ej + cStart, NULL, &coefs));
1759: /* create static point data */
1760: for (PetscInt qj = 0; qj < Nq; qj++, outer_ipidx++) {
1761: const PetscInt gidx = outer_ipidx;
1762: const PetscReal *invJ = &invJe[qj * dim * dim];
1763: ww[gidx] = detJj[qj] * quadWeights[qj];
1764: if (dim == 2) ww[gidx] *= vj[qj * dim + 0]; /* cylindrical coordinate, w/o 2pi */
1765: // get xx, yy, zz
1766: if (ctx->use_energy_tensor_trick) {
1767: double refSpaceDer[3], eGradPhi[3];
1768: const PetscReal *const DD = Tf[0]->T[1];
1769: const PetscReal *Dq = &DD[qj * Nb * dim];
1770: for (PetscInt d = 0; d < 3; ++d) refSpaceDer[d] = eGradPhi[d] = 0.0;
1771: for (PetscInt b = 0; b < Nb; ++b) {
1772: for (PetscInt d = 0; d < dim; ++d) refSpaceDer[d] += Dq[b * dim + d] * PetscRealPart(coefs[b]);
1773: }
1774: xx[gidx] = 1e10;
1775: if (ctx->use_relativistic_corrections) {
1776: double dg2_c2 = 0;
1777: //for (PetscInt d = 0; d < dim; ++d) refSpaceDer[d] *= c02;
1778: for (PetscInt d = 0; d < dim; ++d) dg2_c2 += PetscSqr(refSpaceDer[d]);
1779: dg2_c2 *= (double)c02;
1780: if (dg2_c2 >= .999) {
1781: xx[gidx] = vj[qj * dim + 0]; /* coordinate */
1782: yy[gidx] = vj[qj * dim + 1];
1783: if (dim == 3) zz[gidx] = vj[qj * dim + 2];
1784: 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]));
1785: } else {
1786: PetscReal fact = c02 / PetscSqrtReal(1. - dg2_c2);
1787: for (PetscInt d = 0; d < dim; ++d) refSpaceDer[d] *= fact;
1788: // could test with other point u' that (grad - grad') * U (refSpaceDer, refSpaceDer') == 0
1789: }
1790: }
1791: if (xx[gidx] == 1e10) {
1792: for (PetscInt d = 0; d < dim; ++d) {
1793: for (PetscInt e = 0; e < dim; ++e) eGradPhi[d] += invJ[e * dim + d] * refSpaceDer[e];
1794: }
1795: xx[gidx] = eGradPhi[0];
1796: yy[gidx] = eGradPhi[1];
1797: if (dim == 3) zz[gidx] = eGradPhi[2];
1798: }
1799: } else {
1800: xx[gidx] = vj[qj * dim + 0]; /* coordinate */
1801: yy[gidx] = vj[qj * dim + 1];
1802: if (dim == 3) zz[gidx] = vj[qj * dim + 2];
1803: }
1804: } /* q */
1805: if (ctx->use_energy_tensor_trick) PetscCall(DMPlexVecRestoreClosure(dmEnergy, e_section, v2_2, ej + cStart, NULL, &coefs));
1806: } /* ej */
1807: if (ctx->use_energy_tensor_trick) {
1808: PetscCall(DMRestoreLocalVector(dmEnergy, &v2_2));
1809: PetscCall(DMDestroy(&dmEnergy));
1810: }
1811: } /* grid */
1812: if (ctx->use_energy_tensor_trick) PetscCall(PetscFEDestroy(&fe));
1813: /* cache static data */
1814: if (ctx->deviceType == LANDAU_KOKKOS) {
1815: #if defined(PETSC_HAVE_KOKKOS)
1816: 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));
1817: /* free */
1818: PetscCall(PetscFree4(ww, xx, yy, invJ_a));
1819: if (dim == 3) PetscCall(PetscFree(zz));
1820: #else
1821: SETERRQ(ctx->comm, PETSC_ERR_ARG_WRONG, "-landau_device_type kokkos not built");
1822: #endif
1823: } else { /* CPU version, just copy in, only use part */
1824: 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 ?
1825: ctx->SData_d.w = (void *)ww;
1826: ctx->SData_d.x = (void *)xx;
1827: ctx->SData_d.y = (void *)yy;
1828: ctx->SData_d.z = (void *)zz;
1829: ctx->SData_d.invJ = (void *)invJ_a;
1830: 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));
1831: for (PetscInt ii = 0; ii < ctx->num_species; ii++) {
1832: nu_alpha_p[ii] = nu_alpha[ii];
1833: nu_beta_p[ii] = nu_beta[ii];
1834: invMass_p[ii] = invMass[ii];
1835: }
1836: ctx->SData_d.alpha = (void *)nu_alpha_p;
1837: ctx->SData_d.beta = (void *)nu_beta_p;
1838: ctx->SData_d.invMass = (void *)invMass_p;
1839: ctx->SData_d.lambdas = (void *)lambdas_p;
1840: for (PetscInt grid = 0; grid < LANDAU_MAX_GRIDS; grid++) {
1841: PetscReal (*lambdas)[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS] = (PetscReal (*)[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS])ctx->SData_d.lambdas;
1842: for (PetscInt gridj = 0; gridj < LANDAU_MAX_GRIDS; gridj++) (*lambdas)[grid][gridj] = ctx->lambdas[grid][gridj];
1843: }
1844: }
1845: PetscCall(PetscLogEventEnd(ctx->events[7], 0, 0, 0, 0));
1846: } // initialize
1847: PetscFunctionReturn(PETSC_SUCCESS);
1848: }
1850: /* < v, u > */
1851: 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[])
1852: {
1853: g0[0] = 1.;
1854: }
1856: /* < v, u > */
1857: 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[])
1858: {
1859: static double ttt = 1e-12;
1860: g0[0] = ttt++;
1861: }
1863: /* < v, u > */
1864: 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[])
1865: {
1866: g0[0] = 2. * PETSC_PI * x[0];
1867: }
1869: /*
1870: LandauCreateJacobianMatrix - creates ctx->J with without real data. Hard to keep sparse.
1871: - Like DMPlexLandauCreateMassMatrix. Should remove one and combine
1872: - has old support for field major ordering
1873: */
1874: static PetscErrorCode LandauCreateJacobianMatrix(MPI_Comm comm, Vec X, IS grid_batch_is_inv[LANDAU_MAX_GRIDS], LandauCtx *ctx)
1875: {
1876: PetscInt *idxs = NULL;
1877: Mat subM[LANDAU_MAX_GRIDS];
1879: PetscFunctionBegin;
1880: if (!ctx->gpu_assembly) { /* we need GPU object with GPU assembly */
1881: PetscFunctionReturn(PETSC_SUCCESS);
1882: }
1883: // get the RCM for this grid to separate out species into blocks -- create 'idxs' & 'ctx->batch_is' -- not used
1884: if (ctx->gpu_assembly && ctx->jacobian_field_major_order) PetscCall(PetscMalloc1(ctx->mat_offset[ctx->num_grids] * ctx->batch_sz, &idxs));
1885: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1886: const PetscInt *values, n = ctx->mat_offset[grid + 1] - ctx->mat_offset[grid];
1887: Mat gMat;
1888: DM massDM;
1889: PetscDS prob;
1890: Vec tvec;
1891: // get "mass" matrix for reordering
1892: PetscCall(DMClone(ctx->plex[grid], &massDM));
1893: PetscCall(DMCopyFields(ctx->plex[grid], PETSC_DETERMINE, PETSC_DETERMINE, massDM));
1894: PetscCall(DMCreateDS(massDM));
1895: PetscCall(DMGetDS(massDM, &prob));
1896: 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));
1897: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only")); // this trick is need to both sparsify the matrix and avoid runtime error
1898: PetscCall(DMCreateMatrix(massDM, &gMat));
1899: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only false"));
1900: PetscCall(MatSetOption(gMat, MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE));
1901: PetscCall(MatSetOption(gMat, MAT_IGNORE_ZERO_ENTRIES, PETSC_TRUE));
1902: PetscCall(DMCreateLocalVector(ctx->plex[grid], &tvec));
1903: PetscCall(DMPlexSNESComputeJacobianFEM(massDM, tvec, gMat, gMat, ctx));
1904: PetscCall(MatViewFromOptions(gMat, NULL, "-dm_landau_reorder_mat_view"));
1905: PetscCall(DMDestroy(&massDM));
1906: PetscCall(VecDestroy(&tvec));
1907: subM[grid] = gMat;
1908: if (ctx->gpu_assembly && ctx->jacobian_field_major_order) {
1909: MatOrderingType rtype = MATORDERINGRCM;
1910: IS isrow, isicol;
1911: PetscCall(MatGetOrdering(gMat, rtype, &isrow, &isicol));
1912: PetscCall(ISInvertPermutation(isrow, PETSC_DECIDE, &grid_batch_is_inv[grid]));
1913: PetscCall(ISGetIndices(isrow, &values));
1914: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // add batch size DMs for this species grid
1915: #if !defined(LANDAU_SPECIES_MAJOR)
1916: PetscInt N = ctx->mat_offset[ctx->num_grids], n0 = ctx->mat_offset[grid] + b_id * N;
1917: for (PetscInt ii = 0; ii < n; ++ii) idxs[n0 + ii] = values[ii] + n0;
1918: #else
1919: PetscInt n0 = ctx->mat_offset[grid] * ctx->batch_sz + b_id * n;
1920: for (PetscInt ii = 0; ii < n; ++ii) idxs[n0 + ii] = values[ii] + n0;
1921: #endif
1922: }
1923: PetscCall(ISRestoreIndices(isrow, &values));
1924: PetscCall(ISDestroy(&isrow));
1925: PetscCall(ISDestroy(&isicol));
1926: }
1927: }
1928: 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));
1929: // get a block matrix
1930: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1931: Mat B = subM[grid];
1932: PetscInt nloc, nzl, *colbuf, row, COL_BF_SIZE = 1024;
1933: PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf));
1934: PetscCall(MatGetSize(B, &nloc, NULL));
1935: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
1936: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
1937: const PetscInt *cols;
1938: const PetscScalar *vals;
1939: for (PetscInt i = 0; i < nloc; i++) {
1940: PetscCall(MatGetRow(B, i, &nzl, NULL, NULL));
1941: if (nzl > COL_BF_SIZE) {
1942: PetscCall(PetscFree(colbuf));
1943: 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));
1944: COL_BF_SIZE = nzl;
1945: PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf));
1946: }
1947: PetscCall(MatGetRow(B, i, &nzl, &cols, &vals));
1948: for (PetscInt j = 0; j < nzl; j++) colbuf[j] = cols[j] + moffset;
1949: row = i + moffset;
1950: PetscCall(MatSetValues(ctx->J, 1, &row, nzl, colbuf, vals, INSERT_VALUES));
1951: PetscCall(MatRestoreRow(B, i, &nzl, &cols, &vals));
1952: }
1953: }
1954: PetscCall(PetscFree(colbuf));
1955: }
1956: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(MatDestroy(&subM[grid]));
1957: PetscCall(MatAssemblyBegin(ctx->J, MAT_FINAL_ASSEMBLY));
1958: PetscCall(MatAssemblyEnd(ctx->J, MAT_FINAL_ASSEMBLY));
1960: // debug
1961: PetscCall(MatViewFromOptions(ctx->J, NULL, "-dm_landau_mat_view"));
1962: if (ctx->gpu_assembly && ctx->jacobian_field_major_order) {
1963: Mat mat_block_order;
1964: PetscCall(MatCreateSubMatrix(ctx->J, ctx->batch_is, ctx->batch_is, MAT_INITIAL_MATRIX, &mat_block_order)); // use MatPermute
1965: PetscCall(MatViewFromOptions(mat_block_order, NULL, "-dm_landau_mat_view"));
1966: PetscCall(MatDestroy(&mat_block_order));
1967: PetscCall(VecScatterCreate(X, ctx->batch_is, X, NULL, &ctx->plex_batch));
1968: PetscCall(VecDuplicate(X, &ctx->work_vec));
1969: }
1970: PetscFunctionReturn(PETSC_SUCCESS);
1971: }
1973: static void LandauSphereMapping(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar f[])
1974: {
1975: PetscReal u_max = 0, u_norm = 0, scale, square_inner_radius = PetscRealPart(constants[0]), square_radius = PetscRealPart(constants[1]);
1976: PetscInt d;
1978: for (d = 0; d < dim; ++d) {
1979: PetscReal val = PetscAbsReal(PetscRealPart(u[d]));
1980: if (val > u_max) u_max = val;
1981: u_norm += PetscRealPart(u[d]) * PetscRealPart(u[d]);
1982: }
1983: u_norm = PetscSqrtReal(u_norm);
1985: if (u_max < square_inner_radius) {
1986: for (d = 0; d < dim; ++d) f[d] = u[d];
1987: return;
1988: }
1990: /*
1991: A outer cube has corners at |u| = square_radius.
1992: u_1 is the intersection of the ray with the outer cube face.
1993: R_max = square_radius * sqrt(3) is radius of sphere we want points on outer cube mapped to.
1994: u_0 is the intersection of the ray with the inner cube face.
1995: The cube has corners at |u| = square_inner_radius.
1996: scale to point linearly between u_0 and u_1 so that a point on the inner face does not move, and a point on the outer face moves to the sphere.
1997: */
1998: if (u_max > square_radius + 1e-5) (void)PetscPrintf(PETSC_COMM_SELF, "Error: Point outside outer radius: u_max %g > %g\n", (double)u_max, (double)square_radius);
1999: /* if (PetscAbsReal(u_max - square_inner_radius) < 1e-5 || PetscAbsReal(u_max - square_radius) < 1e-5) {
2000: (void)PetscPrintf(PETSC_COMM_SELF, "Warning: Point near corner of inner and outer cube: u_max %g, inner %g, outer %g\n", (double)u_max, (double)square_inner_radius, (double)square_radius);
2001: } */
2002: {
2003: PetscReal u_0_norm = u_norm * square_inner_radius / u_max;
2004: PetscReal R_max = square_radius * PetscSqrtReal((PetscReal)dim);
2005: PetscReal t = (u_max - square_inner_radius) / (square_radius - square_inner_radius);
2006: PetscReal rho_prime = (1.0 - t) * u_0_norm + t * R_max;
2007: scale = rho_prime / u_norm;
2008: }
2009: for (d = 0; d < dim; ++d) f[d] = u[d] * scale;
2010: }
2012: static PetscErrorCode LandauSphereMesh(DM dm, PetscReal inner, PetscReal radius)
2013: {
2014: DM cdm;
2015: PetscDS cds;
2016: PetscScalar consts[2];
2018: PetscFunctionBegin;
2019: consts[0] = inner;
2020: consts[1] = radius;
2021: PetscCall(DMGetCoordinateDM(dm, &cdm));
2022: PetscCall(DMGetDS(cdm, &cds));
2023: PetscCall(PetscDSSetConstants(cds, 2, consts));
2024: PetscCall(DMPlexRemapGeometry(dm, 0.0, LandauSphereMapping));
2025: PetscFunctionReturn(PETSC_SUCCESS);
2026: }
2028: PetscErrorCode DMPlexLandauCreateMassMatrix(DM pack, Mat *Amat);
2030: /*@C
2031: DMPlexLandauCreateVelocitySpace - Create a `DMPLEX` velocity space mesh
2033: Collective
2035: Input Parameters:
2036: + comm - The MPI communicator
2037: . dim - velocity space dimension (2 for axisymmetric, 3 for full 3X + 3V solver)
2038: - prefix - prefix for options (not tested)
2040: Output Parameters:
2041: + pack - The `DM` object representing the mesh
2042: . X - A vector (user destroys)
2043: - J - Optional matrix (object destroys)
2045: Level: beginner
2047: .seealso: `DMPlexCreate()`, `DMPlexLandauDestroyVelocitySpace()`
2048: @*/
2049: PetscErrorCode DMPlexLandauCreateVelocitySpace(MPI_Comm comm, PetscInt dim, const char prefix[], Vec *X, Mat *J, DM *pack)
2050: {
2051: LandauCtx *ctx;
2052: Vec Xsub[LANDAU_MAX_GRIDS];
2053: IS grid_batch_is_inv[LANDAU_MAX_GRIDS];
2055: PetscFunctionBegin;
2056: PetscCheck(dim == 2 || dim == 3, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Only 2D and 3D supported");
2057: PetscCheck(LANDAU_DIM == dim, PETSC_COMM_SELF, PETSC_ERR_PLIB, "dim %" PetscInt_FMT " != LANDAU_DIM %d", dim, LANDAU_DIM);
2058: PetscCall(PetscNew(&ctx));
2059: ctx->comm = comm; /* used for diagnostics and global errors */
2060: /* process options */
2061: PetscCall(ProcessOptions(ctx, prefix));
2062: if (dim == 2) ctx->use_relativistic_corrections = PETSC_FALSE;
2063: /* Create Mesh */
2064: PetscCall(DMCompositeCreate(PETSC_COMM_SELF, pack));
2065: PetscCall(PetscLogEventBegin(ctx->events[13], 0, 0, 0, 0));
2066: PetscCall(PetscLogEventBegin(ctx->events[15], 0, 0, 0, 0));
2067: PetscCall(LandauDMCreateVMeshes(PETSC_COMM_SELF, dim, prefix, ctx, *pack)); // creates grids (Forest of AMR)
2068: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2069: /* create FEM */
2070: PetscCall(SetupDS(ctx->plex[grid], dim, grid, prefix, ctx));
2071: /* set initial state */
2072: PetscCall(DMCreateGlobalVector(ctx->plex[grid], &Xsub[grid]));
2073: PetscCall(PetscObjectSetName((PetscObject)Xsub[grid], "u_orig"));
2074: /* initial static refinement, no solve */
2075: PetscCall(LandauSetInitialCondition(ctx->plex[grid], Xsub[grid], grid, 0, 1, ctx));
2076: /* forest refinement - forest goes in (if forest), plex comes out */
2077: if (ctx->use_p4est) {
2078: DM plex;
2079: PetscCall(adapt(grid, ctx, &Xsub[grid])); // forest goes in, plex comes out
2080: // convert to plex, all done with this level
2081: PetscCall(DMConvert(ctx->plex[grid], DMPLEX, &plex));
2082: PetscCall(DMDestroy(&ctx->plex[grid]));
2083: ctx->plex[grid] = plex;
2084: } else if (ctx->sphere && dim == 3) {
2085: if (ctx->map_sphere) PetscCall(LandauSphereMesh(ctx->plex[grid], ctx->radius[grid] * ctx->sphere_inner_radius_90degree[grid], ctx->radius[grid]));
2086: PetscCall(LandauSetInitialCondition(ctx->plex[grid], Xsub[grid], grid, 0, 1, ctx));
2087: }
2088: if (grid == 0) {
2089: PetscCall(DMViewFromOptions(ctx->plex[grid], NULL, "-dm_landau_amr_dm_view"));
2090: PetscCall(VecSetOptionsPrefix(Xsub[grid], prefix));
2091: PetscCall(VecViewFromOptions(Xsub[grid], NULL, "-dm_landau_amr_vec_view"));
2092: }
2093: #if !defined(LANDAU_SPECIES_MAJOR)
2094: PetscCall(DMCompositeAddDM(*pack, ctx->plex[grid]));
2095: #else
2096: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // add batch size DMs for this species grid
2097: PetscCall(DMCompositeAddDM(*pack, ctx->plex[grid]));
2098: }
2099: #endif
2100: PetscCall(DMSetApplicationContext(ctx->plex[grid], ctx));
2101: }
2102: #if !defined(LANDAU_SPECIES_MAJOR)
2103: // stack the batched DMs, could do it all here!!! b_id=0
2104: for (PetscInt b_id = 1; b_id < ctx->batch_sz; b_id++) {
2105: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMCompositeAddDM(*pack, ctx->plex[grid]));
2106: }
2107: #endif
2108: // create ctx->mat_offset
2109: ctx->mat_offset[0] = 0;
2110: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2111: PetscInt n;
2112: PetscCall(VecGetLocalSize(Xsub[grid], &n));
2113: ctx->mat_offset[grid + 1] = ctx->mat_offset[grid] + n;
2114: }
2115: // creat DM & Jac
2116: PetscCall(DMSetApplicationContext(*pack, ctx));
2117: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only"));
2118: PetscCall(DMCreateMatrix(*pack, &ctx->J));
2119: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only false"));
2120: PetscCall(MatSetOption(ctx->J, MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE));
2121: PetscCall(MatSetOption(ctx->J, MAT_IGNORE_ZERO_ENTRIES, PETSC_TRUE));
2122: PetscCall(PetscObjectSetName((PetscObject)ctx->J, "Jac"));
2123: // construct initial conditions in X
2124: PetscCall(DMCreateGlobalVector(*pack, X));
2125: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2126: PetscInt n;
2127: PetscCall(VecGetLocalSize(Xsub[grid], &n));
2128: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
2129: PetscScalar const *values;
2130: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
2131: PetscCall(LandauSetInitialCondition(ctx->plex[grid], Xsub[grid], grid, b_id, ctx->batch_sz, ctx));
2132: PetscCall(VecGetArrayRead(Xsub[grid], &values)); // Drop whole grid in Plex ordering
2133: for (PetscInt i = 0, idx = moffset; i < n; i++, idx++) PetscCall(VecSetValue(*X, idx, values[i], INSERT_VALUES));
2134: PetscCall(VecRestoreArrayRead(Xsub[grid], &values));
2135: }
2136: }
2137: // cleanup
2138: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(VecDestroy(&Xsub[grid]));
2139: /* check for correct matrix type */
2140: if (ctx->gpu_assembly) { /* we need GPU object with GPU assembly */
2141: PetscBool flg;
2142: if (ctx->deviceType == LANDAU_KOKKOS) {
2143: PetscCall(PetscObjectTypeCompareAny((PetscObject)ctx->J, &flg, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, MATAIJKOKKOS, ""));
2144: #if defined(PETSC_HAVE_KOKKOS)
2145: 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'");
2146: #else
2147: 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'");
2148: #endif
2149: }
2150: }
2151: PetscCall(PetscLogEventEnd(ctx->events[15], 0, 0, 0, 0));
2153: // create field major ordering
2154: ctx->work_vec = NULL;
2155: ctx->plex_batch = NULL;
2156: ctx->batch_is = NULL;
2157: for (PetscInt i = 0; i < LANDAU_MAX_GRIDS; i++) grid_batch_is_inv[i] = NULL;
2158: PetscCall(PetscLogEventBegin(ctx->events[12], 0, 0, 0, 0));
2159: PetscCall(LandauCreateJacobianMatrix(comm, *X, grid_batch_is_inv, ctx));
2160: PetscCall(PetscLogEventEnd(ctx->events[12], 0, 0, 0, 0));
2162: // create AMR GPU assembly maps and static GPU data
2163: PetscCall(CreateStaticData(dim, grid_batch_is_inv, prefix, ctx));
2165: PetscCall(PetscLogEventEnd(ctx->events[13], 0, 0, 0, 0));
2167: // create mass matrix
2168: PetscCall(DMPlexLandauCreateMassMatrix(*pack, NULL));
2170: if (J) *J = ctx->J;
2172: if (ctx->gpu_assembly && ctx->jacobian_field_major_order) {
2173: PetscContainer container;
2174: // cache ctx for KSP with batch/field major Jacobian ordering -ksp_type gmres/etc -dm_landau_jacobian_field_major_order
2175: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
2176: PetscCall(PetscContainerSetPointer(container, (void *)ctx));
2177: PetscCall(PetscObjectCompose((PetscObject)ctx->J, "LandauCtx", (PetscObject)container));
2178: PetscCall(PetscContainerDestroy(&container));
2179: // batch solvers need to map -- can batch solvers work
2180: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
2181: PetscCall(PetscContainerSetPointer(container, (void *)ctx->plex_batch));
2182: PetscCall(PetscObjectCompose((PetscObject)ctx->J, "plex_batch_is", (PetscObject)container));
2183: PetscCall(PetscContainerDestroy(&container));
2184: }
2185: // for batch solvers
2186: {
2187: PetscContainer container;
2188: PetscInt *pNf;
2189: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
2190: PetscCall(PetscMalloc1(sizeof(*pNf), &pNf));
2191: *pNf = ctx->batch_sz;
2192: PetscCall(PetscContainerSetPointer(container, (void *)pNf));
2193: PetscCall(PetscContainerSetCtxDestroy(container, PetscCtxDestroyDefault));
2194: PetscCall(PetscObjectCompose((PetscObject)ctx->J, "batch size", (PetscObject)container));
2195: PetscCall(PetscContainerDestroy(&container));
2196: }
2197: PetscFunctionReturn(PETSC_SUCCESS);
2198: }
2200: /*@C
2201: DMPlexLandauAccess - Access to the distribution function with user callback
2203: Collective
2205: Input Parameters:
2206: + pack - the `DMCOMPOSITE`
2207: . func - call back function
2208: - user_ctx - user context
2210: Input/Output Parameter:
2211: . X - Vector to data to
2213: Level: advanced
2215: .seealso: `DMPlexLandauCreateVelocitySpace()`
2216: @*/
2217: PetscErrorCode DMPlexLandauAccess(DM pack, Vec X, PetscErrorCode (*func)(DM, Vec, PetscInt, PetscInt, PetscInt, void *), void *user_ctx)
2218: {
2219: LandauCtx *ctx;
2221: PetscFunctionBegin;
2222: PetscCall(DMGetApplicationContext(pack, &ctx)); // uses ctx->num_grids; ctx->plex[grid]; ctx->batch_sz; ctx->mat_offset
2223: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2224: PetscInt dim, n;
2225: PetscCall(DMGetDimension(pack, &dim));
2226: for (PetscInt sp = ctx->species_offset[grid], i0 = 0; sp < ctx->species_offset[grid + 1]; sp++, i0++) {
2227: Vec vec;
2228: PetscInt vf[1] = {i0};
2229: IS vis;
2230: DM vdm;
2231: PetscCall(DMCreateSubDM(ctx->plex[grid], 1, vf, &vis, &vdm));
2232: PetscCall(DMSetApplicationContext(vdm, ctx)); // the user might want this
2233: PetscCall(DMCreateGlobalVector(vdm, &vec));
2234: PetscCall(VecGetSize(vec, &n));
2235: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
2236: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
2237: PetscCall(VecZeroEntries(vec));
2238: /* Add your data with 'dm' for species 'sp' to 'vec' */
2239: PetscCall(func(vdm, vec, i0, grid, b_id, user_ctx));
2240: /* add to global */
2241: PetscScalar const *values;
2242: const PetscInt *offsets;
2243: PetscCall(VecGetArrayRead(vec, &values));
2244: PetscCall(ISGetIndices(vis, &offsets));
2245: for (PetscInt i = 0; i < n; i++) PetscCall(VecSetValue(X, moffset + offsets[i], values[i], ADD_VALUES));
2246: PetscCall(VecRestoreArrayRead(vec, &values));
2247: PetscCall(ISRestoreIndices(vis, &offsets));
2248: } // batch
2249: PetscCall(VecDestroy(&vec));
2250: PetscCall(ISDestroy(&vis));
2251: PetscCall(DMDestroy(&vdm));
2252: }
2253: } // grid
2254: PetscFunctionReturn(PETSC_SUCCESS);
2255: }
2257: /*@
2258: DMPlexLandauDestroyVelocitySpace - Destroy a `DMPLEX` velocity space mesh
2260: Collective
2262: Input/Output Parameters:
2263: . dm - the `DM` to destroy
2265: Level: beginner
2267: .seealso: `DMPlexLandauCreateVelocitySpace()`
2268: @*/
2269: PetscErrorCode DMPlexLandauDestroyVelocitySpace(DM *dm)
2270: {
2271: LandauCtx *ctx;
2273: PetscFunctionBegin;
2274: PetscCall(DMGetApplicationContext(*dm, &ctx));
2275: PetscCall(MatDestroy(&ctx->M));
2276: PetscCall(MatDestroy(&ctx->J));
2277: for (PetscInt ii = 0; ii < ctx->num_species; ii++) PetscCall(PetscFEDestroy(&ctx->fe[ii]));
2278: PetscCall(ISDestroy(&ctx->batch_is));
2279: PetscCall(VecDestroy(&ctx->work_vec));
2280: PetscCall(VecScatterDestroy(&ctx->plex_batch));
2281: if (ctx->deviceType == LANDAU_KOKKOS) {
2282: #if defined(PETSC_HAVE_KOKKOS)
2283: PetscCall(LandauKokkosStaticDataClear(&ctx->SData_d));
2284: #else
2285: SETERRQ(ctx->comm, PETSC_ERR_ARG_WRONG, "-landau_device_type %s not built", "kokkos");
2286: #endif
2287: } else {
2288: if (ctx->SData_d.x) { /* in a CPU run */
2289: 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;
2290: 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;
2291: PetscCall(PetscFree4(ww, xx, yy, invJ));
2292: if (zz) PetscCall(PetscFree(zz));
2293: if (coo_elem_offsets) PetscCall(PetscFree3(coo_elem_offsets, coo_elem_fullNb, coo_elem_point_offsets)); // could be NULL
2294: PetscCall(PetscFree4(ctx->SData_d.alpha, ctx->SData_d.beta, ctx->SData_d.invMass, ctx->SData_d.lambdas));
2295: }
2296: }
2298: if (ctx->times[LANDAU_MATRIX_TOTAL] > 0) { // OMP timings
2299: PetscCall(PetscPrintf(ctx->comm, "TSStep N 1.0 %10.3e\n", ctx->times[LANDAU_EX2_TSSOLVE]));
2300: 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));
2301: PetscCall(PetscPrintf(ctx->comm, "3: Landau: %10.3e\n", ctx->times[LANDAU_MATRIX_TOTAL]));
2302: PetscCall(PetscPrintf(ctx->comm, "Landau Jacobian %" PetscInt_FMT " 1.0 %10.3e\n", (PetscInt)ctx->times[LANDAU_JACOBIAN_COUNT], ctx->times[LANDAU_JACOBIAN]));
2303: PetscCall(PetscPrintf(ctx->comm, "Landau Operator N 1.0 %10.3e\n", ctx->times[LANDAU_OPERATOR]));
2304: PetscCall(PetscPrintf(ctx->comm, "Landau Mass N 1.0 %10.3e\n", ctx->times[LANDAU_MASS]));
2305: PetscCall(PetscPrintf(ctx->comm, " Jac-f-df (GPU) N 1.0 %10.3e\n", ctx->times[LANDAU_F_DF]));
2306: PetscCall(PetscPrintf(ctx->comm, " Kernel (GPU) N 1.0 %10.3e\n", ctx->times[LANDAU_KERNEL]));
2307: PetscCall(PetscPrintf(ctx->comm, "MatLUFactorNum X 1.0 %10.3e\n", ctx->times[KSP_FACTOR]));
2308: PetscCall(PetscPrintf(ctx->comm, "MatSolve X 1.0 %10.3e\n", ctx->times[KSP_SOLVE]));
2309: }
2310: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMDestroy(&ctx->plex[grid]));
2311: PetscCall(PetscFree(ctx));
2312: PetscCall(DMDestroy(dm));
2313: PetscFunctionReturn(PETSC_SUCCESS);
2314: }
2316: /* < v, ru > */
2317: 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)
2318: {
2319: PetscInt ii = (PetscInt)PetscRealPart(constants[0]);
2320: f0[0] = u[ii];
2321: }
2323: /* < v, ru > */
2324: 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)
2325: {
2326: PetscInt ii = (PetscInt)PetscRealPart(constants[0]), jj = (PetscInt)PetscRealPart(constants[1]);
2327: f0[0] = x[jj] * u[ii]; /* x momentum */
2328: }
2330: 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)
2331: {
2332: PetscInt i, ii = (PetscInt)PetscRealPart(constants[0]);
2333: double tmp1 = 0.;
2334: for (i = 0; i < dim; ++i) tmp1 += x[i] * x[i];
2335: f0[0] = tmp1 * u[ii];
2336: }
2338: static PetscErrorCode gamma_n_f(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf, PetscScalar *u, void *actx)
2339: {
2340: const PetscReal *c2_0_arr = ((PetscReal *)actx);
2341: const PetscReal c02 = c2_0_arr[0];
2343: PetscFunctionBegin;
2344: for (PetscInt s = 0; s < Nf; s++) {
2345: PetscReal tmp1 = 0.;
2346: for (PetscInt i = 0; i < dim; ++i) tmp1 += x[i] * x[i];
2347: #if defined(PETSC_USE_DEBUG)
2348: u[s] = PetscSqrtReal(1. + tmp1 / c02); // u[0] = PetscSqrtReal(1. + xx);
2349: #else
2350: {
2351: PetscReal xx = tmp1 / c02;
2352: u[s] = xx / (PetscSqrtReal(1. + xx) + 1.); // better conditioned = xx/(PetscSqrtReal(1. + xx) + 1.)
2353: }
2354: #endif
2355: }
2356: PetscFunctionReturn(PETSC_SUCCESS);
2357: }
2359: /* < v, ru > */
2360: 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)
2361: {
2362: PetscInt ii = (PetscInt)PetscRealPart(constants[0]);
2363: f0[0] = 2. * PETSC_PI * x[0] * u[ii];
2364: }
2366: /* < v, ru > */
2367: 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)
2368: {
2369: PetscInt ii = (PetscInt)PetscRealPart(constants[0]);
2370: f0[0] = 2. * PETSC_PI * x[0] * x[1] * u[ii];
2371: }
2373: 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)
2374: {
2375: PetscInt ii = (PetscInt)PetscRealPart(constants[0]);
2376: f0[0] = 2. * PETSC_PI * x[0] * (x[0] * x[0] + x[1] * x[1]) * u[ii];
2377: }
2379: /*@
2380: DMPlexLandauPrintNorms - collects moments and prints them
2382: Collective
2384: Input Parameters:
2385: + X - the state
2386: - stepi - current step to print
2388: Level: beginner
2390: .seealso: `DMPlexLandauCreateVelocitySpace()`
2391: @*/
2392: PetscErrorCode DMPlexLandauPrintNorms(Vec X, PetscInt stepi)
2393: {
2394: LandauCtx *ctx;
2395: PetscDS prob;
2396: DM pack;
2397: PetscInt cStart, cEnd, dim, ii, i0, nDMs;
2398: PetscScalar xmomentumtot = 0, ymomentumtot = 0, zmomentumtot = 0, energytot = 0, densitytot = 0, tt[LANDAU_MAX_SPECIES];
2399: PetscScalar xmomentum[LANDAU_MAX_SPECIES], ymomentum[LANDAU_MAX_SPECIES], zmomentum[LANDAU_MAX_SPECIES], energy[LANDAU_MAX_SPECIES], density[LANDAU_MAX_SPECIES];
2400: Vec *globXArray;
2402: PetscFunctionBegin;
2403: PetscCall(VecGetDM(X, &pack));
2404: PetscCheck(pack, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Vector has no DM");
2405: PetscCall(DMGetDimension(pack, &dim));
2406: PetscCheck(dim == 2 || dim == 3, PETSC_COMM_SELF, PETSC_ERR_PLIB, "dim %" PetscInt_FMT " not in [2,3]", dim);
2407: PetscCall(DMGetApplicationContext(pack, &ctx));
2408: PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context");
2409: /* print momentum and energy */
2410: PetscCall(DMCompositeGetNumberDM(pack, &nDMs));
2411: 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);
2412: PetscCall(PetscMalloc(sizeof(*globXArray) * nDMs, &globXArray));
2413: PetscCall(DMCompositeGetAccessArray(pack, X, nDMs, NULL, globXArray));
2414: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2415: Vec Xloc = globXArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)];
2416: PetscCall(DMGetDS(ctx->plex[grid], &prob));
2417: for (ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) {
2418: PetscScalar user[2] = {(PetscScalar)i0, ctx->charges[ii]};
2419: PetscCall(PetscDSSetConstants(prob, 2, user));
2420: if (dim == 2) { /* 2/3X + 3V (cylindrical coordinates) */
2421: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_rden));
2422: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2423: density[ii] = tt[0] * ctx->n_0 * ctx->charges[ii];
2424: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_rmom));
2425: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2426: zmomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii];
2427: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_rv2));
2428: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2429: energy[ii] = tt[0] * 0.5 * ctx->n_0 * ctx->v_0 * ctx->v_0 * ctx->masses[ii];
2430: zmomentumtot += zmomentum[ii];
2431: energytot += energy[ii];
2432: densitytot += density[ii];
2433: 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])));
2434: } else { /* 2/3Xloc + 3V */
2435: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_den));
2436: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2437: density[ii] = tt[0] * ctx->n_0 * ctx->charges[ii];
2438: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_mom));
2439: user[1] = 0;
2440: PetscCall(PetscDSSetConstants(prob, 2, user));
2441: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2442: xmomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii];
2443: user[1] = 1;
2444: PetscCall(PetscDSSetConstants(prob, 2, user));
2445: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2446: ymomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii];
2447: user[1] = 2;
2448: PetscCall(PetscDSSetConstants(prob, 2, user));
2449: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2450: zmomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii];
2451: if (ctx->use_relativistic_corrections) {
2452: /* gamma * M * f */
2453: if (ii == 0 && grid == 0) { // do all at once
2454: Vec Mf, globGamma, *globMfArray, *globGammaArray;
2455: PetscErrorCode (*gammaf[1])(PetscInt, PetscReal, const PetscReal[], PetscInt, PetscScalar[], void *) = {gamma_n_f};
2456: PetscReal *c2_0[1], data[1];
2458: PetscCall(VecDuplicate(X, &globGamma));
2459: PetscCall(VecDuplicate(X, &Mf));
2460: PetscCall(PetscMalloc(sizeof(*globMfArray) * nDMs, &globMfArray));
2461: PetscCall(PetscMalloc(sizeof(*globMfArray) * nDMs, &globGammaArray));
2462: /* M * f */
2463: PetscCall(MatMult(ctx->M, X, Mf));
2464: /* gamma */
2465: PetscCall(DMCompositeGetAccessArray(pack, globGamma, nDMs, NULL, globGammaArray));
2466: 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
2467: Vec v1 = globGammaArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)];
2468: data[0] = PetscSqr(C_0(ctx->v_0));
2469: c2_0[0] = &data[0];
2470: PetscCall(DMProjectFunction(ctx->plex[grid], 0., gammaf, (void **)c2_0, INSERT_ALL_VALUES, v1));
2471: }
2472: PetscCall(DMCompositeRestoreAccessArray(pack, globGamma, nDMs, NULL, globGammaArray));
2473: /* gamma * Mf */
2474: PetscCall(DMCompositeGetAccessArray(pack, globGamma, nDMs, NULL, globGammaArray));
2475: PetscCall(DMCompositeGetAccessArray(pack, Mf, nDMs, NULL, globMfArray));
2476: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { // yes a grid loop in a grid loop to print nice
2477: PetscInt Nf = ctx->species_offset[grid + 1] - ctx->species_offset[grid], N, bs;
2478: Vec Mfsub = globMfArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)], Gsub = globGammaArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)], v1, v2;
2479: // get each component
2480: PetscCall(VecGetSize(Mfsub, &N));
2481: PetscCall(VecCreate(ctx->comm, &v1));
2482: PetscCall(VecSetSizes(v1, PETSC_DECIDE, N / Nf));
2483: PetscCall(VecCreate(ctx->comm, &v2));
2484: PetscCall(VecSetSizes(v2, PETSC_DECIDE, N / Nf));
2485: PetscCall(VecSetFromOptions(v1)); // ???
2486: PetscCall(VecSetFromOptions(v2));
2487: // get each component
2488: PetscCall(VecGetBlockSize(Gsub, &bs));
2489: PetscCheck(bs == Nf, PETSC_COMM_SELF, PETSC_ERR_PLIB, "bs %" PetscInt_FMT " != num_species %" PetscInt_FMT " in Gsub", bs, Nf);
2490: PetscCall(VecGetBlockSize(Mfsub, &bs));
2491: PetscCheck(bs == Nf, PETSC_COMM_SELF, PETSC_ERR_PLIB, "bs %" PetscInt_FMT " != num_species %" PetscInt_FMT, bs, Nf);
2492: for (PetscInt i = 0, ix = ctx->species_offset[grid]; i < Nf; i++, ix++) {
2493: PetscScalar val;
2494: PetscCall(VecStrideGather(Gsub, i, v1, INSERT_VALUES)); // this is not right -- TODO
2495: PetscCall(VecStrideGather(Mfsub, i, v2, INSERT_VALUES));
2496: PetscCall(VecDot(v1, v2, &val));
2497: energy[ix] = PetscRealPart(val) * ctx->n_0 * ctx->v_0 * ctx->v_0 * ctx->masses[ix];
2498: }
2499: PetscCall(VecDestroy(&v1));
2500: PetscCall(VecDestroy(&v2));
2501: } /* grids */
2502: PetscCall(DMCompositeRestoreAccessArray(pack, globGamma, nDMs, NULL, globGammaArray));
2503: PetscCall(DMCompositeRestoreAccessArray(pack, Mf, nDMs, NULL, globMfArray));
2504: PetscCall(PetscFree(globGammaArray));
2505: PetscCall(PetscFree(globMfArray));
2506: PetscCall(VecDestroy(&globGamma));
2507: PetscCall(VecDestroy(&Mf));
2508: }
2509: } else {
2510: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_v2));
2511: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2512: energy[ii] = 0.5 * tt[0] * ctx->n_0 * ctx->v_0 * ctx->v_0 * ctx->masses[ii];
2513: }
2514: 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])));
2515: xmomentumtot += xmomentum[ii];
2516: ymomentumtot += ymomentum[ii];
2517: zmomentumtot += zmomentum[ii];
2518: energytot += energy[ii];
2519: densitytot += density[ii];
2520: }
2521: if (ctx->num_species > 1) PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
2522: }
2523: }
2524: PetscCall(DMCompositeRestoreAccessArray(pack, X, nDMs, NULL, globXArray));
2525: PetscCall(PetscFree(globXArray));
2526: /* totals */
2527: PetscCall(DMPlexGetHeightStratum(ctx->plex[0], 0, &cStart, &cEnd));
2528: if (ctx->num_species > 1) {
2529: if (dim == 2) {
2530: 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),
2531: (double)(ctx->masses[1] / ctx->masses[0]), cEnd - cStart));
2532: } else {
2533: 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),
2534: (double)(ctx->masses[1] / ctx->masses[0]), cEnd - cStart));
2535: }
2536: } else PetscCall(PetscPrintf(PETSC_COMM_WORLD, " -- %" PetscInt_FMT " cells", cEnd - cStart));
2537: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
2538: PetscFunctionReturn(PETSC_SUCCESS);
2539: }
2541: /*@
2542: DMPlexLandauCreateMassMatrix - Create mass matrix for Landau in Plex space (not field major order of Jacobian)
2543: - puts mass matrix into ctx->M
2545: Collective
2547: Input Parameter:
2548: . pack - the `DM` object. Puts matrix in Landau context M field
2550: Output Parameter:
2551: . Amat - The mass matrix (optional), mass matrix is added to the `DM` context
2553: Level: beginner
2555: .seealso: `DMPlexLandauCreateVelocitySpace()`
2556: @*/
2557: PetscErrorCode DMPlexLandauCreateMassMatrix(DM pack, Mat *Amat)
2558: {
2559: DM mass_pack, massDM[LANDAU_MAX_GRIDS];
2560: PetscDS prob;
2561: PetscInt ii, dim, N1 = 1, N2;
2562: LandauCtx *ctx;
2563: Mat packM, subM[LANDAU_MAX_GRIDS];
2565: PetscFunctionBegin;
2567: if (Amat) PetscAssertPointer(Amat, 2);
2568: PetscCall(DMGetApplicationContext(pack, &ctx));
2569: PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context");
2570: PetscCall(PetscLogEventBegin(ctx->events[14], 0, 0, 0, 0));
2571: PetscCall(DMGetDimension(pack, &dim));
2572: PetscCall(DMCompositeCreate(PetscObjectComm((PetscObject)pack), &mass_pack));
2573: /* create pack mass matrix */
2574: for (PetscInt grid = 0, ix = 0; grid < ctx->num_grids; grid++) {
2575: PetscCall(DMClone(ctx->plex[grid], &massDM[grid]));
2576: PetscCall(DMCopyFields(ctx->plex[grid], PETSC_DETERMINE, PETSC_DETERMINE, massDM[grid]));
2577: PetscCall(DMCreateDS(massDM[grid]));
2578: PetscCall(DMGetDS(massDM[grid], &prob));
2579: for (ix = 0, ii = ctx->species_offset[grid]; ii < ctx->species_offset[grid + 1]; ii++, ix++) {
2580: if (dim == 3) PetscCall(PetscDSSetJacobian(prob, ix, ix, g0_1, NULL, NULL, NULL));
2581: else PetscCall(PetscDSSetJacobian(prob, ix, ix, g0_r, NULL, NULL, NULL));
2582: }
2583: #if !defined(LANDAU_SPECIES_MAJOR)
2584: PetscCall(DMCompositeAddDM(mass_pack, massDM[grid]));
2585: #else
2586: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // add batch size DMs for this species grid
2587: PetscCall(DMCompositeAddDM(mass_pack, massDM[grid]));
2588: }
2589: #endif
2590: PetscCall(DMCreateMatrix(massDM[grid], &subM[grid]));
2591: }
2592: #if !defined(LANDAU_SPECIES_MAJOR)
2593: // stack the batched DMs
2594: for (PetscInt b_id = 1; b_id < ctx->batch_sz; b_id++) {
2595: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMCompositeAddDM(mass_pack, massDM[grid]));
2596: }
2597: #endif
2598: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only"));
2599: PetscCall(DMCreateMatrix(mass_pack, &packM));
2600: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only false"));
2601: PetscCall(MatSetOption(packM, MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE));
2602: PetscCall(MatSetOption(packM, MAT_IGNORE_ZERO_ENTRIES, PETSC_TRUE));
2603: PetscCall(DMDestroy(&mass_pack));
2604: /* make mass matrix for each block */
2605: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2606: Vec locX;
2607: DM plex = massDM[grid];
2608: PetscCall(DMGetLocalVector(plex, &locX));
2609: /* Mass matrix is independent of the input, so no need to fill locX */
2610: PetscCall(DMPlexSNESComputeJacobianFEM(plex, locX, subM[grid], subM[grid], ctx));
2611: PetscCall(DMRestoreLocalVector(plex, &locX));
2612: PetscCall(DMDestroy(&massDM[grid]));
2613: }
2614: PetscCall(MatGetSize(ctx->J, &N1, NULL));
2615: PetscCall(MatGetSize(packM, &N2, NULL));
2616: PetscCheck(N1 == N2, PetscObjectComm((PetscObject)pack), PETSC_ERR_PLIB, "Incorrect matrix sizes: |Jacobian| = %" PetscInt_FMT ", |Mass|=%" PetscInt_FMT, N1, N2);
2617: /* assemble block diagonals */
2618: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2619: Mat B = subM[grid];
2620: PetscInt nloc, nzl, *colbuf, COL_BF_SIZE = 1024, row;
2621: PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf));
2622: PetscCall(MatGetSize(B, &nloc, NULL));
2623: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
2624: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
2625: const PetscInt *cols;
2626: const PetscScalar *vals;
2627: for (PetscInt i = 0; i < nloc; i++) {
2628: PetscCall(MatGetRow(B, i, &nzl, NULL, NULL));
2629: if (nzl > COL_BF_SIZE) {
2630: PetscCall(PetscFree(colbuf));
2631: PetscCall(PetscInfo(pack, "Realloc buffer %" PetscInt_FMT " to %" PetscInt_FMT " (row size %" PetscInt_FMT ") \n", COL_BF_SIZE, 2 * COL_BF_SIZE, nzl));
2632: COL_BF_SIZE = nzl;
2633: PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf));
2634: }
2635: PetscCall(MatGetRow(B, i, &nzl, &cols, &vals));
2636: for (PetscInt j = 0; j < nzl; j++) colbuf[j] = cols[j] + moffset;
2637: row = i + moffset;
2638: PetscCall(MatSetValues(packM, 1, &row, nzl, colbuf, vals, INSERT_VALUES));
2639: PetscCall(MatRestoreRow(B, i, &nzl, &cols, &vals));
2640: }
2641: }
2642: PetscCall(PetscFree(colbuf));
2643: }
2644: // cleanup
2645: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(MatDestroy(&subM[grid]));
2646: PetscCall(MatAssemblyBegin(packM, MAT_FINAL_ASSEMBLY));
2647: PetscCall(MatAssemblyEnd(packM, MAT_FINAL_ASSEMBLY));
2648: PetscCall(PetscObjectSetName((PetscObject)packM, "mass"));
2649: PetscCall(MatViewFromOptions(packM, NULL, "-dm_landau_mass_view"));
2650: ctx->M = packM;
2651: if (Amat) *Amat = packM;
2652: PetscCall(PetscLogEventEnd(ctx->events[14], 0, 0, 0, 0));
2653: PetscFunctionReturn(PETSC_SUCCESS);
2654: }
2656: /*@
2657: DMPlexLandauIFunction - `TS` residual calculation, confusingly this computes the Jacobian w/o mass
2659: Collective
2661: Input Parameters:
2662: + ts - The time stepping context
2663: . time_dummy - current time (not used)
2664: . X - Current state
2665: . X_t - Time derivative of current state
2666: - actx - Landau context
2668: Output Parameter:
2669: . F - The residual
2671: Level: beginner
2673: .seealso: `DMPlexLandauCreateVelocitySpace()`, `DMPlexLandauIJacobian()`
2674: @*/
2675: PetscErrorCode DMPlexLandauIFunction(TS ts, PetscReal time_dummy, Vec X, Vec X_t, Vec F, void *actx)
2676: {
2677: LandauCtx *ctx = (LandauCtx *)actx;
2678: PetscInt dim;
2679: DM pack;
2680: #if defined(PETSC_HAVE_THREADSAFETY)
2681: double starttime, endtime;
2682: #endif
2683: PetscObjectState state;
2685: PetscFunctionBegin;
2686: PetscCall(TSGetDM(ts, &pack));
2687: PetscCall(DMGetApplicationContext(pack, &ctx));
2688: PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context");
2689: if (ctx->stage) PetscCall(PetscLogStagePush(ctx->stage));
2690: PetscCall(PetscLogEventBegin(ctx->events[11], 0, 0, 0, 0));
2691: PetscCall(PetscLogEventBegin(ctx->events[0], 0, 0, 0, 0));
2692: #if defined(PETSC_HAVE_THREADSAFETY)
2693: starttime = MPI_Wtime();
2694: #endif
2695: PetscCall(DMGetDimension(pack, &dim));
2696: PetscCall(PetscObjectStateGet((PetscObject)ctx->J, &state));
2697: if (state != ctx->norm_state) {
2698: PetscCall(MatZeroEntries(ctx->J));
2699: PetscCall(LandauFormJacobian_Internal(X, ctx->J, dim, 0.0, (void *)ctx));
2700: PetscCall(MatViewFromOptions(ctx->J, NULL, "-dm_landau_jacobian_view"));
2701: PetscCall(PetscObjectStateGet((PetscObject)ctx->J, &state));
2702: ctx->norm_state = state;
2703: } else {
2704: PetscCall(PetscInfo(ts, "WARNING Skip forming Jacobian, has not changed %" PetscInt64_FMT "\n", state));
2705: }
2706: /* mat vec for op */
2707: PetscCall(MatMult(ctx->J, X, F)); /* C*f */
2708: /* add time term */
2709: if (X_t) PetscCall(MatMultAdd(ctx->M, X_t, F, F));
2710: #if defined(PETSC_HAVE_THREADSAFETY)
2711: if (ctx->stage) {
2712: endtime = MPI_Wtime();
2713: ctx->times[LANDAU_OPERATOR] += (endtime - starttime);
2714: ctx->times[LANDAU_JACOBIAN] += (endtime - starttime);
2715: ctx->times[LANDAU_MATRIX_TOTAL] += (endtime - starttime);
2716: ctx->times[LANDAU_JACOBIAN_COUNT] += 1;
2717: }
2718: #endif
2719: PetscCall(PetscLogEventEnd(ctx->events[0], 0, 0, 0, 0));
2720: PetscCall(PetscLogEventEnd(ctx->events[11], 0, 0, 0, 0));
2721: if (ctx->stage) PetscCall(PetscLogStagePop());
2722: PetscFunctionReturn(PETSC_SUCCESS);
2723: }
2725: /*@
2726: DMPlexLandauIJacobian - `TS` Jacobian construction, confusingly this adds mass
2728: Collective
2730: Input Parameters:
2731: + ts - The time stepping context
2732: . time_dummy - current time (not used)
2733: . X - Current state
2734: . U_tdummy - Time derivative of current state (not used)
2735: . shift - shift for du/dt term
2736: - actx - Landau context
2738: Output Parameters:
2739: + Amat - Jacobian
2740: - Pmat - same as Amat
2742: Level: beginner
2744: .seealso: `DMPlexLandauCreateVelocitySpace()`, `DMPlexLandauIFunction()`
2745: @*/
2746: PetscErrorCode DMPlexLandauIJacobian(TS ts, PetscReal time_dummy, Vec X, Vec U_tdummy, PetscReal shift, Mat Amat, Mat Pmat, void *actx)
2747: {
2748: LandauCtx *ctx = NULL;
2749: PetscInt dim;
2750: DM pack;
2751: #if defined(PETSC_HAVE_THREADSAFETY)
2752: double starttime, endtime;
2753: #endif
2754: PetscObjectState state;
2756: PetscFunctionBegin;
2757: PetscCall(TSGetDM(ts, &pack));
2758: PetscCall(DMGetApplicationContext(pack, &ctx));
2759: PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context");
2760: PetscCheck(Amat == Pmat && Amat == ctx->J, ctx->comm, PETSC_ERR_PLIB, "Amat!=Pmat || Amat!=ctx->J");
2761: PetscCall(DMGetDimension(pack, &dim));
2762: /* get collision Jacobian into A */
2763: if (ctx->stage) PetscCall(PetscLogStagePush(ctx->stage));
2764: PetscCall(PetscLogEventBegin(ctx->events[11], 0, 0, 0, 0));
2765: PetscCall(PetscLogEventBegin(ctx->events[9], 0, 0, 0, 0));
2766: #if defined(PETSC_HAVE_THREADSAFETY)
2767: starttime = MPI_Wtime();
2768: #endif
2769: PetscCheck(shift != 0.0, ctx->comm, PETSC_ERR_PLIB, "zero shift");
2770: PetscCall(PetscObjectStateGet((PetscObject)ctx->J, &state));
2771: PetscCheck(state == ctx->norm_state, ctx->comm, PETSC_ERR_PLIB, "wrong state, %" PetscInt64_FMT " %" PetscInt64_FMT, ctx->norm_state, state);
2772: if (!ctx->use_matrix_mass) {
2773: PetscCall(LandauFormJacobian_Internal(X, ctx->J, dim, shift, (void *)ctx));
2774: } else { /* add mass */
2775: PetscCall(MatAXPY(Pmat, shift, ctx->M, SAME_NONZERO_PATTERN));
2776: }
2777: #if defined(PETSC_HAVE_THREADSAFETY)
2778: if (ctx->stage) {
2779: endtime = MPI_Wtime();
2780: ctx->times[LANDAU_OPERATOR] += (endtime - starttime);
2781: ctx->times[LANDAU_MASS] += (endtime - starttime);
2782: ctx->times[LANDAU_MATRIX_TOTAL] += (endtime - starttime);
2783: }
2784: #endif
2785: PetscCall(PetscLogEventEnd(ctx->events[9], 0, 0, 0, 0));
2786: PetscCall(PetscLogEventEnd(ctx->events[11], 0, 0, 0, 0));
2787: if (ctx->stage) PetscCall(PetscLogStagePop());
2788: PetscFunctionReturn(PETSC_SUCCESS);
2789: }